Simultaneous Denoising and Baseline Correction of Microplate Raman Spectra Using a Dual-Branch U‑Net

Abstract

We present a dual-branch U-Net architecture for simultaneous baseline correction and denoising of Raman spectra. The network features a shared encoder that diverges into two specialized decoding heads for the Raman signal and for the baseline. The two heads are coupled with a cross-branch gating mechanism. The model can cross-confirm the peaks by comparing the recovered Raman signal with the baseline-corrected spectrum. Moreover, the model supports quantitative analysis by integrating the deep Raman-head reconstruction, I _deep, expressed in the detector’s intensity units. The model was trained entirely using a custom synthetic data engine explicitly designed to emulate automated HTS acquisitions from microplates. Comprehensive validation demonstrates robust peak recovery on synthetic spectra with signal-to-noise ratios (SNRs) as low as 5. The model recovers the Raman signal from highly noisy glycerol samples and from moderately noisy adenine sulfate samples. Quantitative analysis is conducted on guanine by integrating I _deep at a deep decoder layer. Transfer is tested on l-tyrosine, which is chemically unrelated to the training distribution.

1. Introduction

Raman spectroscopy has emerged as a powerful analytical technique across diverse scientific disciplines, including pharmaceutical science, materials science, biomedical research, and environmental science. Its widespread adoption is driven by the technique’s ability to provide a label-free, nondestructive molecular fingerprint of a sample. This is because Raman scattering relies on the inelastic scattering of light where the energy shift of scattered photons directly corresponds to the vibrational modes of target molecules. Despite this analytical power, the throughput of Raman data acquisition continues to lag behind that of other spectroscopic methods. While 96- and 384-well microplates are ubiquitous standards for high-throughput screening (HTS) in fields like genomics and drug discovery, their integration into Raman spectroscopy remains severely restricted. A major barrier is the complexity of automated screening. These challenges are typically either mechanical, related to the precision motion systems required for well-by-well screening, or optical, stemming from the difficulties of simultaneous excitation and signal collection. Consequently, commercial high-throughput Raman systems require substantial capital investments, rendering them inaccessible to many research laboratories.

While automated platforms like RamanBot , resolve the mechanical and optical bottlenecks of high-throughput screening, a fundamental physical barrier remains: the severe SNR challenge inherent to Raman scattering. Only about one in every 10⁶ incident photons scatters inelastically. Consequently, the inherently weak Raman signal is frequently obscured by two dominant artifacts: high-frequency thermal and shot noise from the detector, and a broad, intense low-frequency baseline primarily caused by sample autofluorescence. This issue is exacerbated in HTS applications, as standard microplates are manufactured from polymers such as polystyrene or polypropylene. These materials generate massive fluorescence signatures that can easily overwhelm the weak signal of the sample within the well. The scale of this problem is apparent in the background of a water-filled microplate well, in which the fluorescence background dominates the residual Raman signal. For microplate measurements a dedicated treatment is a prerequisite for recovering any analyte signature at all, which is the reason our architecture devotes a separate decoding branch to it. Historically, extracting the true molecular signal has relied on conventional mathematical algorithms. Techniques such as Savitzky-Golay filtering or wavelet transforms are standard for high-frequency denoising, while Asymmetric Least Squares (ALS) and Adaptive Iteratively Reweighted Penalized Least Squares (airPLS) are standard for baseline removal. However, these conventional methods suffer from critical drawbacks. They require fine-tuning of parameters for each sample type, and they routinely fail or distort peak shapes when applied to spectra with extremely low SNRs. Additionally, applying these techniques consecutively leads to loss of peaks or false peaks due to noise. These distortions make tasks such as classification or quantification highly sensitive to the cleanliness of the extracted Raman peaks. To overcome the limitations of classical algorithms, there has been a recent shift toward applying deep learning to Raman spectral processing. Convolutional Neural Networks (CNNs) and Residual Network architectures have been successfully deployed to automate baseline correction and noise reduction. − Specifically for denoising, an encoder-decoder network was proposed to enhance the SNR of noisy data, enabling 10-fold faster acquisition. Similarly, a generative adversarial CNN was introduced to reduce acquisition time by an order of magnitude for single-cell analysis, and a U-Net architecture based on self-supervised learning was constructed for the generic denoising of Raman signals. For baseline removal, a residual U-Net significantly improved the limit of detection (LOD) of Rh6G to 10^–8 M. Addressing both baseline removal and spectrum denoising simultaneously is often achieved by cascading models, where one model handles the baseline and the other tackles the noise. ,, Although successful, this cascaded approach is fundamentally vulnerable to missing subtle peaks or distorting peak intensities. Because the networks operate independently without a shared representational connection, errors generated in the first stage can easily propagate to the second, causing an irreversible loss of chemical information. , Moreover, cascaded architectures introduce significant computational and optimization inefficiencies. Deploying two independent models also requires two full forward passes. Two forward passes do not imply higher latency, because the cost of a pass depends on the network that executes it. Furthermore, because the networks are optimized as disjointed entities, they are unable to leverage shared feature representations. This isolates the gradient flow, preventing the models from synergistically learning the physical relationship between the low-frequency fluorescence and the sparse Raman scattering, which often leads to suboptimal global convergence.

The limitations of treating the two degradations separately have long been recognized. Liu et al. coupled baseline estimation and denoising in a single variational model, on the grounds that a baseline estimated from a strongly noisy spectrum is itself unreliable. That formulation is a classical optimization: it carries no learned prior over peak shape, and its regularization parameters must be chosen for each spectrum.

Joint variational models and cascaded learned pipelines both remain in use. Huang et al. combined channel and spatial attention with a cascaded ResUNet for the reconstruction and unmixing of mixed microplastic spectra; because the stages are applied in sequence, detail suppressed by the first is unavailable to the second. Liao et al. proposed MGD-CNN, a convolutional baseline estimator followed by an enhancement network optimized under a single objective; its baseline estimator is supervised against the output of a morphological–Gaussian algorithm, so the baseline it can learn is bounded by that of the reference method. Our approach differs from both in deriving its interbranch guidance from an explicitly supervised baseline estimate, and in using the internal state of the network as a quantitative readout.

Building on the idea of treating baseline and noise jointly, we present a dual-branch U-Net architecture for simultaneous Raman baseline correction and denoising. This network features a shared encoder, which avoids, by construction, the sequential information loss of cascaded approaches. The encoder is constructed from Runge–Kutta (RK4)-inspired residual blocks, which further reduce encoder error. Following the shared bottleneck, the network branches into two distinct decoding heads: one dedicated to extracting the clean Raman signal and the other to recovering the low-frequency baseline. Fundamentally, this dual-branch approach operates as a hard-parameter sharing Multi-Task Learning (MTL) framework, enabling the network to leverage shared feature representations rather than relying on isolated gradient flows. This joint optimization encourages the encoder to learn generalized features. These features are capable of decoupling the superimposed baseline and Raman signals without destructive interference.

Moreover, this approach provides a built-in validation mechanism, allowing the derived signal (raw spectrum – generated baseline) to be directly compared against the network’s generated Raman signal. Furthermore, the architecture facilitates continuous interhead communication via a cross-branch gate, a baseline-guided spatial gating module related to the additive attention gates of Oktay et al. This gate provides spatial “hints” regarding the location of potential peaks by evaluating the differences between the intermediate encoder layers and the baseline’s intermediate layers. By concurrently updating the shared encoder, the two heads work synergistically to reduce the likelihood that subtle spectral features are missed. Deep supervision is applied to the deep and middle layers of the decoder to ensure the alignment of the features. , The network is trained on a robust synthetic data set generated by a custom Raman spectrum engine, which was specifically developed to accurately mimic the high-throughput output of RamanBot acquiring samples directly from microplates. To ensure real-world applicability, the model’s performance is subsequently validated on glycerol and adenine sulfate acquired directly from microplates via the RamanBot platform, and transfer to an unrelated analyte class is tested on l-tyrosine. Finally, quantitative and acquisition-time analysis is performed on guanine samples by integrating I _deep at the deep Raman decoder layer.

2. Experimental Setup

For data acquisition, we used the RamanBot , a previously developed automated screening platform. As illustrated in Figure , the system repurposes the motion gantry of a 3D printer, replacing the standard extrusion head with a custom-designed, 3D-printed optical module. We have previously demonstrated that the RamanBot can autonomously screen microplates by sequentially translating between wells according to a preprogrammed trajectory. This high-throughput capability makes it an ideal platform for rapidly acquiring the large-scale Raman data sets required for machine learning applications.

1.

(a) A schematic diagram of the Raman setup. (b) The 3D printed head that mounts all the optics needed to excite wells sequentially.

The internal optical configuration of the head is designed for efficient signal excitation and collection. Excitation light is delivered to the module via a multimode (MM) fiber and collimated by a 0.22 numerical aperture (NA) package. The beam passes through a notch filter to suppress laser sidebands before a 0.4 NA lens focuses the light onto the sample well. The backscattered Raman signal is collected by the same lens and reflected by a dichroic mirror toward the collection path. The signal then passes through a long-pass filter to further reject Rayleigh scattering. Finally, a second 0.22 NA lens focuses the filtered Raman signal into a fiber bundle, which routes the collected light to an Andor HoloSpec spectrometer for detection.

Figure shows the raw background spectrum captured from one well filled with 330 μL of water. The spectrum was obtained under 1 min acquisition time. This figure shows how complex the background signal of a microplate is. This configuration (a well containing 330 μL of water and no analyte) is the same one used to acquire the empirical background library from which the synthetic training baselines are subsequently constructed. Figure shows a single baseline of that library. Water is used because it isolates the contribution of the microplate and the solvent from that of any analyte, so that the resulting library characterizes the instrument and the consumable, not a particular sample. The dashed red line represents the baseline computed using the airPLS technique. airPLS is used as a familiar and widely adopted reference for visualizing the magnitude of the background relative to the residual signal. The technique is used later as the classical comparator against which the proposed model is evaluated. The inset shows the baseline-corrected Raman spectrum of the microwell. In the first 700 cm^–1, an oscillatory pattern is observed that can be attributed to the etalon effect. The pattern arises from multiple reflections between the microwell bottom and the liquid surface. This oscillation is an optical interference effect, not a fluorescence background, and is consequently not something a baseline-estimation algorithm can be expected to remove. Its persistence in the corrected residual is precisely why we favor a model trained on background spectra acquired from this instrument and this consumable over a generic smoothness-based baseline estimate.

2.

Raw Raman background spectrum and baseline correction obtained from a 96-well plate filled with 330 μL of water. A dashed red line denotes the computed baseline fitted across the raw data. The inset displays the baseline-corrected signal.

3. Methodology

3.1. Synthetic Raman Data Engine

For the deep learning model to achieve optimal performance, it requires training on a substantial volume of high-quality data. Acquiring such a large data set experimentally is prohibitively time-consuming, and it would lack a perfectly clean ground truth. Previous models have been trained on synthetic spectra. ,, Those generators were not specific to this instrument or to the intended sample class. Consequently, we developed a synthetic data engine capable of generating extensive spectral data sets that accurately simulate the performance of our experimental setup. The generator models a measured spectrum, X(ν), as a linear superposition of the Raman signal S(ν), the fluorescence baseline B(ν), and the noise N(ν), where ν represents the Raman shift:

$$X(\nu )=S(\nu )+B(\nu )+N(\nu )$$ 1

3.1.1. Signal Modeling

Because the samples in the microplate are predominantly liquids, their Raman peaks fundamentally exhibit a Lorentzian profile. However, instrumental broadening introduced by the optical setup and the CCD detector imposes a Gaussian contribution, resulting in faster-decaying tails. Therefore, to accurately capture the convolution of natural line widths and the instrumental response, we model the Raman peaks using a pseudo-Voigt profile, which is a linear combination of Gaussian (G) and Lorentzian (L) functions:

$$S(\nu )=\sum _{k=1}^{N_p}{A}_k·[\eta L(\nu ;{\nu }_k,{\sigma }_k)+(1-\eta )G(\nu ;{\nu }_k,{\sigma }_k)]$$ 2

where N _p is the number of peaks, η is a weighting parameter that controls the profile shape, A _k is the peak amplitude, and σ _k is the peak width.

3.1.2. Baseline

Inspired by the approach in ref , which generates new baselines through the superposition of experimental measurements, we collected 1,000 background spectra from a microplate well containing 330 μL of water using acquisition times ranging from 1 to 60 s. A synthetic baseline is first generated via linear interpolation between two randomly selected experimental spectra. We extend this method by superimposing a low-frequency spline curve, P _n , constructed from n random points distributed across the Raman shift range. This addition simulates physical phenomena that alter the baseline shape, such as sample evaporation or slight optical misfocusing. The final generated baseline can thus be expressed as

$$B(\nu )=(1-r)[\alpha B_l+(1-\alpha )B_m]+r{P}_n$$ 3

where α controls the blending ratio between the two empirical baselines, B _l and B _m , while r dictates the relative contribution of the synthetic spline to the final baseline.

3.1.3. Noise

Noise added to the synthetic Raman signal is usually expressed mathematically. , However, to accurately model the inherent noise of our experimental setup, we acquired 1,000 dark current spectra using various integration times with the spectrometer shutter closed. Additionally, we collected 1,000 blank spectra across different integration times by focusing the excitation laser 20 cm away from a black microplate to capture ambient and background system noise. To construct a synthetic noise profile, we randomly combine these two noise sources and scale the resulting distribution to achieve the target SNR before superimposing it onto the generated Raman signal.

Figure illustrates three synthetic Raman spectra generated using the proposed pipeline at distinct SNR levels (25, 15, and 5). Each spectrum is synthesized with a randomized combination of peak shifts, amplitudes, and line widths. Furthermore, the simulated baseline variations across the three generated spectra are clearly observable, reflecting the diverse background conditions accurately captured by our model.

3.

Generation of synthetic Raman spectra under varying noise conditions. The synthetic data were constructed by varying peak parameters (number, width, amplitude) and superimposing a fluorescent baseline. The blue solid line represents the total noisy signal, the red dashed line shows the generated baseline, and the green solid line depicts the pure Raman signal. The panels display representative spectra at SNRs of (a) 25, (b) 15, and (c) 5.

3.1.4. Training Data Set

Table summarizes the parameters used to generate the training data sets. Each spectrum contains between 5 and 14 peaks, and a fixed fraction of 5% of training spectra are generated signal-free, so that the network is explicitly exposed to backgrounds containing no analyte. The Raman peak shifts were selected within the range of 341 to 1861 cm^–1, that is, no closer than 20 points to either end of the detector, to prevent spectral truncation at the boundaries of the Raman shift range. The shape parameter, η, was chosen to vary between 0.5 and 1.0; this favors a Lorentzian distribution, which is characteristic of the liquid samples used in this study. The peak-width parameter is drawn on a logarithmic scale between 3 and 30 points, which favors sharp peaks and corresponds to a full width at half-maximum between 5.5 and 55.4 cm^–1 at the dispersion of this instrument. The spline contribution, r, is restricted to the interval [0, 0.3]. This low range reflects the experimental setup involving dark 96-well microplates, where only minimal baseline variations are expected. We used a random number of spline points in the range [3,8] where we restricted the first and last point to be at the edges of the Raman shift range while the locations of the remaining points were picked randomly. The blended empirical baseline is additionally scaled by a factor drawn between 2 and 10, so that the ratio of background to Raman signal varies across the data set. Finally, given the inherently weak nature of Raman signals, we generated a data set with SNRs ranging from 4 to 60 for training and from 4 to 26 for the validation set. These values were sampled on a logarithmic scale to ensure the model is robustly trained across both low- and high-SNR regimes.

1. Training Dataset Parameters.

Metric	Value Range
Spectral axis (cm–1)	[294.9, 1888.2]
Points per spectrum	864
Spectral dispersion (cm–1/point)	1.85
Number of peaks per spectrum	[5, 14]
Peak shift ν k cm–1	[341, 1861]
Peak shape η	[0.5, 1.0]
Peak amplitude A k	[0.1, 1.0]
Peak width parameter, log-uniform (points)	[3, 30]
resulting σ k (cm–1)	[2.8, 27.7]
resulting fwhm (cm–1)	[5.5, 55.4]
Baseline superposition α	[0, 1.0]
Baseline amplitude	[2, 10]
Spline contribution r	[0, 0.3]
Number of spline points n	[3, 8]
Signal-free (blank) fraction, training	0.05
SNR, training (log-uniform)	[4, 60]
SNR, validation (log-uniform)	[4, 26]

^aValues are those used by the generator that produced the reported model. Ranges are sampled uniformly unless stated otherwise.

3.2. Model Architecture

We propose a dual-branch U-Net architecture, illustrated in Figure a, designed specifically to decouple complex baseline artifacts from weak Raman scattering. The network accepts a 1D Raman spectrum of 864 points following laser line removal. The first residual block expands the single channel to 64 channels. The network then uses a shared encoder composed of three down-sampling blocks that progressively compress the spatial sequence to a length of 108 at the bottleneck while increasing the feature depth to 512 channels. Each down-block is responsible for spatial reduction with advanced ordinary differential equation (ODE) inspired feature extraction. It consists of a 1D convolution layer with a kernel size of 3 and a stride of 2, which efficiently halves the spatial resolution of the input sequence while maintaining the initial channel depth. The downscaled tensor is immediately normalized using group normalization (8 groups) to stabilize the learning process without relying on batch statistics, followed by a standard LeakyReLU activation. Finally, the tensor is passed into an RK4-inspired residual block. ,, This expands the feature space to the targeted channels while refining the representations using a fourth-order RK4-style residual integration, offering enhanced gradient flow and stability compared to standard residual blocks. We quantify the contribution of this design in the Supporting Information (Section S4.2). Replacing the RK4 block with a standard preactivation residual block increases the validation mean absolute error by 41.8% and the cosine loss by 88.5%, while reducing the parameter count by only 10%. Specifically, the block evaluates the tensor through an internal sequence, F(z), comprising two 1D convolutions (kernel size 3, padding 1) that are each preceded by group normalization and a ReLU activation. The feature representation is then updated using the explicit RK4 integration step: $z_{out}=z_{in}+\frac{h}{6}(k_1+2k_2+2k_3+k_4)$ , where h = 1.0 and the k terms represent the four intermediate state evaluations. To account for the low-frequency nature of the expected macroscopic fluorescence, the 512 bottleneck channels are divided asymmetrically: 496 channels for the Raman head and 16 channels for the baseline head. We deliberately assign more channels to the Raman head to accommodate the greater complexity of the Raman component. Strict latent space disentanglement between the Raman and baseline channels is enforced through an orthogonal auxiliary loss. This penalizes overlap between the baseline features and the Raman features. From this latent representation, the architecture branches into two specialized, jointly trained decoders. The 16 baseline channels and 496 Raman channels are independently up-converted using up-blocks to reconstruct the baseline and Raman signals, respectively. Each up-block is designed to restore spatial resolution while fusing deep semantic features with lower-level spatial details from the encoder. It comprises a learnable up-convolution layer with a kernel size of 2 and a stride of 2, doubling the spatial resolution while halving the channel depth, with padding used to guarantee perfect spatial alignment. The corresponding encoder and decoder tensors are concatenated along the channel dimension to bridge their context, then fed into a preactivated residual block to fuse and refine the features. This residual block consists of two consecutive sequences of group normalization, a ReLU activation, and a 1D convolution (kernel size 3, padding 1). The raw input is added directly to this residual output via an identity skip connection.

4.

Dual-Branch U-Net architecture. (a) The full network. The Baseline Head (bottom branch) estimates the background. This estimate is subtracted from the encoder features to create a “Hint” map, which drives the cross-branch gate in the Raman Head (top branch), amplifying weak peaks. Blocks labeled RK4 employ fourth-order Runge–Kutta residual integration. Tensor dimensions are given as sequence length × channels. (b) The cross-branch gate, labeled G in (a).

The Raman branch communicates with the baseline branch via a cross-branch gate, shown in Figure b. We use the term cross-branch to denote that the gating signal originates in the opposing decoder branch. The module is a spatial gate, not an attention mechanism as in Vaswani et al. It produces a single-channel multiplicative mask over the spectral axis. However, it differs from additive attention gates of Attention U-Net in that the gating signal is derived from the residual between the encoder features and the baseline decoder features, rather than from a coarser-scale copy of the same branch. At each decoder stage, this gate forms its guidance signal by subtracting the intermediate baseline features from the corresponding high-resolution encoder features. This difference, which we refer to as the “hint”, is processed via a 1D convolution to compress the channel dimension. This subtraction is not a residual in the physical sense of S = X – B. Measurements reported in the Supporting Information (Section S1.3) show that corresponding channels of the two tensors are essentially uncorrelated. The results of replacing the subtraction with a concatenation can be found in the Supporting Information. Withholding the baseline features from the gate altogether degrades the mean absolute error by 5.9%. The operative quantity is therefore the information the gate receives about the network’s current estimate of the background. The subtraction acts as an economical means of presenting that information, not a physical operation. To prevent high-frequency noise from passing through, a deep-semantic gating signal from the previous layer is linearly interpolated and concatenated with the hint signal. A subsequent convolution block and a sigmoid activation function create a continuous spatial mask used to gate the hint. Consequently, physical evidence is propagated in proportion to its corroboration by the deep semantic context. To ensure this gating mechanism operates on physically accurate spatial drafts, we incorporate deep supervision , at the intermediate decoder layers (highlighted with dotted lines in Figure ). These auxiliary losses serve as macroscopic physical anchors. By applying average pooling to downsample the high-resolution ground truth to match the spatial resolution of the intermediate decoders, we penalize deviations from local energy conservation. This encourages the deep layers to map the total photon count and broad geometry of the signals. As a result, the final high-resolution convolutions sharpen features that are already present at low resolution, without reproducing high-frequency shot noise. To suppress residual noise in regions lacking genuine chemical signatures, a final spatial squelch gate is positioned at the terminal end of the Raman decoder. Constructed from a spatial 1D convolution paired with a sigmoid function, the gate generates a continuous 1D probability mask across the sequence length. By applying this mask to the raw signal prediction, the gate acts as an adaptive filter that squelches interpeak noise floors to zero while opening over the recovered peaks to preserve their pseudo-Voigt profiles. A critical advantage of this dual-branch architecture is the utilization of a shared encoder and a unified bottleneck, fundamentally distinguishing it from conventional cascaded networks. In traditional sequential approacheswhere one network models the baseline and a subsequent network performs signal denoisingthe system suffers from severe error propagation; if the first network erroneously suppresses a weak Raman peak, the subsequent network cannot recover the lost chemical information. By employing a shared encoder, our network constructs a holistic latent representation of the entire optical signal at the bottleneck, preserving the physical context of both the baseline background and the Raman peaks. This allows the deep Raman layer to hold a bulk representation of the Raman photons, which is subsequently used for quantitative analysis. Furthermore, branching from a single bottleneck enables joint optimization through multitask learning. During training, gradients from both the baseline and Raman heads concurrently update the shared encoder, acting as mutual regularizers. This forces the encoder to learn highly robust, generalized feature extractors that effectively decouple the superimposed signals without destructive interference, and a single forward pass yields both the baseline and the Raman estimate.

3.3. Quantitative Readout at the Deep Raman Layer

The spectrometer reports a photon count at each of its 864 detector pixels, so the measured spectrum X(ν) is expressed directly in photon counts. Two scalings are applied before inference. Each spectrum is min–max normalized to the interval [0,1], and the Raman targets used during training are multiplied by a fixed factor of 10 to improve the numerical conditioning of weak signals. The deep auxiliary head of the Raman decoder outputs a 216-point reconstruction ŝ ^deep of the Raman component at one-quarter of the input sampling density, that is, one sample per 7.38 cm^–1. Reversing both scalings returns this reconstruction to the intensity units of the measurement, and summing it along the spectral axis gives the integral of the recovered Raman component. Because the deep head is sampled at one-quarter of the input density and the sum is not rescaled for that, this integral is approximately a quarter of the corresponding full-resolution integral, as Section S8.2 of the Supporting Information quantifies:

$$I_{\mathrm{deep}}=\frac{X_{max}-X_{min}}{10}\sum _{i=1}^{216}{ŝ}_i^{\mathrm{deep}}$$ 4

where X _max and X _min are the maximum and minimum of the raw measured spectrum. I _deep is the integral of the network’s own reconstruction of the Raman component, expressed in the intensity units reported by the detector rather than as a calibrated photon flux. It is proportional to the true number of Raman photons incident on the detector up to the gain and quantum efficiency of the CCD, which are not independently calibrated in this work, and it inherits the normalization span X _max–X _min of the spectrum from which it is computed.

The readout is taken at the deep layer, and not at the network output, because the auxiliary loss at that layer is computed against an average-pooled ground truth, which constrains the integrated intensity over each pooled interval and not the position of individual points. The deep layer is therefore optimized for the conservation of total signal, whereas the final high-resolution output is additionally shaped by the terminal squelch gate, which is optimized for the suppression of the interpeak noise floor and not for the preservation of integrated area.

Sampling at one-quarter of the input density requires that narrow bands remain adequately represented. Sweeps of peak width and of subsample peak position, reported in Section S7, show no evidence of aliasing. The recovered integral stays centered on the conserved value down to 0.75 samples per full width at half-maximum, and the same sweeps on the full-resolution output reproduce that result. Two effects appear at the extremes of the range. Broad bands are under-reported, by up to 18% at a full width at half-maximum of 55 cm^–1, because the curvature penalty regularizing the baseline ceases to distinguish a band from the background as the band broadens; the missing area is recovered in the baseline branch (see Section S7.3). The narrowest bands remain unbiased but lose precision, with the standard deviation across background realizations reaching 45% at a full width at half-maximum of 5.5 cm^–1. The quantitative claims of this work are accordingly bounded to bands between approximately 9 and 37 cm^–1, over which the recovered integral remains within 8% of the conserved value.

3.4. Loss Function

Our total loss function, ${\mathcal{L}}_{total}$ , is formulated as a weighted sum of the signal loss $({\mathcal{L}}_s)$ , the baseline loss $({\mathcal{L}}_b)$ , and an orthogonality loss $({\mathcal{L}}_{ortho})$ , governed by their respective hyperparameters w _s , w _b , and w _ortho . These loss components are formally defined as follows:

$$\begin{array}{ll}{\mathcal{L}}_{total}& =w_b{\mathcal{L}}_b+w_s{\mathcal{L}}_s+w_{ortho}{\mathcal{L}}_{ortho}\\ {\mathcal{L}}_b& ={\mathcal{L}}_{mse}+w_{tv1}{\mathcal{L}}_{tv1}+w_{tv2}{\mathcal{L}}_{tv2}+{\mathcal{L}}_b^{aux}\\ {\mathcal{L}}_s& ={\mathcal{L}}_{cos}+{\mathcal{L}}_{L1}+w_{shape}{\mathcal{L}}_{shape}+{\mathcal{L}}_s^{aux}\end{array}$$ 5

Here, the baseline loss ${\mathcal{L}}_b$ is composed of a data-fidelity term ${\mathcal{L}}_{mse}$ representing the mean squared error (MSE) between the predicted and ground-truth baselinesalongside an auxiliary baseline loss, ${\mathcal{L}}_b^{aux}$ . To ensure the predicted baseline reflects the physically slow-varying nature of a typical background, we introduce first- and second-order derivative penalties ( ${\mathcal{L}}_{tv1}$ and ${\mathcal{L}}_{tv2}$ ), scaled by w _tv1 and w _tv2, respectively. These regularization terms explicitly penalize abrupt changes in slope and curvature, thereby enforcing spatial smoothness on the generated baseline.

The signal loss, ${\mathcal{L}}_s$ , integrates four components: a cosine loss ( ${\mathcal{L}}_{cos}=1-\cos \theta$ , where cosθ is the cosine similarity between the predicted and ground-truth Raman vectors; we use cosine similarity for the quantity that approaches unity for a perfect match, and cosine loss for the training term that approaches zero), a mean absolute error (MAE) for amplitude fidelity $({\mathcal{L}}_{L1})$ , a shape loss $({\mathcal{L}}_{shape})$ representing the MAE between the first derivatives of the predicted and ground-truth Raman signals, and an auxiliary signal loss $({\mathcal{L}}_s^{aux})$ . The orthogonality loss, ${\mathcal{L}}_{ortho}$ , is defined as the squared Frobenius norm of the cross-correlation matrix between the baseline latent channels and the signal latent channels, encouraging optimal feature separation. The auxiliary losses, ${\mathcal{L}}_b^{aux}$ and ${\mathcal{L}}_s^{aux}$ , act as intermediate structural anchors that force the network to learn the macroscopic physics of a sample before attempting to resolve high-frequency details. They are expressed as

$$\begin{array}{ll}{\mathcal{L}}_b^{aux}& =w_{deep}{\mathcal{L}}_b^{deep}+w_{mid}{\mathcal{L}}_b^{mid}\\ {\mathcal{L}}_s^{aux}& =w_{deep}{\mathcal{L}}_s^{deep}+w_{mid}{\mathcal{L}}_s^{mid}\end{array}$$ 6

where ${\mathcal{L}}_b^{deep}$ and ${\mathcal{L}}_s^{deep}$ denote the baseline and signal losses at the deep layers, respectively, while ${\mathcal{L}}_b^{mid}$ and ${\mathcal{L}}_s^{mid}$ represent the corresponding losses at the intermediate layers. The hyperparameters w _deep and w _mid govern the relative contribution of these deep and middle layers during training. Specifically, the auxiliary baseline loss, ${\mathcal{L}}_b^{aux}$ , is computed as the mean squared error (MSE) between the target baseline curve and the intermediate predictions of the baseline branch. Conversely, the auxiliary signal loss, ${\mathcal{L}}_s^{aux}$ , combines the MAE and cosine similarity losses; importantly, it is computed between the ground-truth Raman peaks and the low-resolution spatial maps extracted directly from the deep bottleneck layers of the signal branch.

Table summarizes the hyperparameter values configured for the loss function. The signal and baseline losses are assigned equal weights (w _s and w _b ), despite the primary focus being the extraction of the inherently weak Raman signal. This balanced weighting is strategic: an accurate baseline predictionwhich the network learns readily due to the background’s high amplitude and low-frequency variationsprovides a highly reliable hint value for the gating mechanism, ultimately boosting overall model performance. Furthermore, the first- and second-order derivative weights (w _tv1 and w _tv2) are both set to 0.2 to enforce baseline smoothness while maintaining a tight data fit to the input spectrum. Finally, the shape penalty weight (w _shape ) is set to 0.1 to rigorously preserve the pseudo-Voigt profile of the detected Raman peaks.

2. Loss Function Parameters.

Metric	Value
Signal, baseline weight w s , w b	20, 20
Signal shape weight w shape	0.1
Baseline roughness weight w tv1	0.2
Baseline curvature weight w tv2	0.2
Auxiliary deep weight w deep	1.0
Auxiliary middle weight w mid	2.0
Bottleneck orthogonality w ortho	0.5

4. Training

The network was trained using the Adam optimizer to minimize the composite multitask loss function with initial learning rate of 1 × 10^–4. Training was conducted using a batch size of 64 on a dynamically generated data set of synthetic spectra. Furthermore, all convolutional weights were initialized using Kaiming normal initialization. Training was conducted on a single NVIDIA RTX 5070 Ti GPU. To ensure stable convergence and prevent gradient explosion across the deep architecture, gradient clipping was enforced with a maximum norm of 1.0. The training process was dynamically modulated using a performance-driven learning rate annealing strategy. Rather than monitoring the total composite loss, the scheduler strictly tracked the validation MAE of the isolated Raman signal prediction, ensuring that learning rate adjustments were driven solely by the network’s quantitative accuracy. If the validation MAE failed to improve for 15 consecutive epochs, the learning rate was reduced by a factor of 0.5 (down to a minimum threshold of 1 × 10^–8) to facilitate fine weight adjustments near local minima. The validation set comprised 2,000 samples with SNRs logarithmically distributed between 4 and 26. To prevent overfitting and capture the optimal network state, an early stopping mechanism with a patience of 100 epochs was implemented. This mechanism continuously monitored the validation MAE, saving the model weights only when a new global minimum was achieved. The model was trained for a maximum of 1,000 epochs, with early stopping triggered at epoch 759. Upon completion, the training loop automatically restored these best-fit weights, ensuring the final evaluated model represents peak quantitative performance rather than a potentially overfitted state. Figure a illustrates the progression of the training and validation losses across epochs. The apparent elevation of the training loss relative to the validation loss is a direct result of the assigned hyperparameters and the inclusion of the orthogonality and auxiliary penalties; the validation loss, by contrast, comprises only the unweighted baseline and signal errors. Figure b details the evolution of the MAE, cosine, and shape losses for the Raman branch. The shape loss drops rapidly and smoothly to approximately 0.007, while the cosine loss converges more gradually to 0.013, indicating that the network readily learns the overall spectral morphology. Conversely, the MAE decreases slowly to ∼0.02, reflecting the inherent difficulty of accurately reconstructing absolute amplitude valuesa challenge stemming from the intrinsically weak nature of Raman peaks relative to the dominant background, particularly at low SNRs. Finally, the orthogonality and auxiliary signal losses are presented in Figure c. The orthogonality loss drops sharply to ∼0.0003. This minimal value confirms a highly effective disentanglement of the baseline and Raman latent spaces, and its rapid convergence validates that the asymmetrical channel allocation (16 for the baseline head and 496 for the Raman head) provides sufficient representational capacity. Furthermore, the stabilization of the auxiliary signal loss at ∼0.07 suggests that the deep and middle layers of the Raman branch successfully preserve the total photon count and structural integrity of the ground truth, effectively managing the lower spatial resolutions inherent to these intermediate layers.

5.

Training loss performance over the epochs broken down into (a) overall training and validation loss, (b) validation metrics (MAE, COS, and shape loss), and (c) orthogonality and auxiliary loss. COS denotes the cosine loss ${\mathcal{L}}_{cos}=1-\cos \theta$ defined in eq . It is plotted here as a loss so that it decreases toward zero.

5. Results and Discussion

5.1. Validation with Synthetic Data

To systematically quantify the model’s robustness against varying noise levels, we evaluated its predictive accuracy across a continuous range of SNRs from 5 to 25. For this comprehensive analysis, a dedicated test set comprising 500 synthetic spectra was generated for each SNR increment. Figure illustrates the average Mean Squared Error (MSE), Mean Absolute Error (MAE), and cosine similarity computed across these data sets. As anticipated, the network’s performance improves monotonically as the SNR increases. The top and middle panels demonstrate that both the MSE and MAE follow similar decay trajectories. At extreme noise levels (SNR = 5), the absolute amplitude errors are at their highest (MSE ≈0.005, MAE ≈0.031), reflecting the inherent difficulty of extracting precise intensity values from heavy background interference. However, as the signal quality improves toward SNR 25, these error metrics decrease steadily, reaching minimum values of approximately 0.0015 and 0.020, respectively. This trend highlights the model’s strong quantitative amplitude recovery, particularly as moderate-to-low noise conditions are reached. The bottom panel tracks the cosine similarity, which serves as a primary metric for overall spectral shape fidelity. Even under the most severe noise conditions tested (SNR = 5), the model maintains a strong baseline cosine similarity of approximately 0.980, indicating that the core morphological features of the Raman peaks are largely preserved despite the heavy noise. As the SNR increases, this metric climbs rapidly before asymptotically approaching a near-perfect score of ∼0.995 at SNR 25. Collectively, these quantitative trends indicate that the proposed network delivers reliable qualitative (shape) and quantitative (amplitude) spectral reconstruction across the range of conditions tested.

6.

Average MSE, MAE, and cosine similarity computed at different SNR values. A total of 500 spectra were used for each SNR.

To qualitatively evaluate the model’s performance across varying noise levels, we synthesized three representative Raman spectra at SNRs of 5, 15, and 25. Figure a provides a detailed breakdown of the most challenging case (SNR = 5). This spectrum was constructed with five peaks located at 400, 602, 810, 1300, and 1400 cm^–1, featuring varying line widths. Notably, the broad features at 1300 and 1400 cm^–1 intentionally overlap; this configuration tests the network’s capacity to resolve convoluted peaks and accurately differentiate broad Raman scattering from the underlying background baseline. Additionally, a synthetic cosmic ray spike was injected into a signal-free region at 1680 cm^–1 to assess the model’s robustness against false-positive detection. The panels in Figure a systematically deconstruct the model’s output. The top panel displays the raw synthetic spectrum alongside the model-generated baseline (dotted line), demonstrating a smooth, accurate fit to the slowly varying background trend. The middle panel contrasts the ground-truth Raman signal with the baseline-corrected spectrum. Finally, the bottom panel compares the model’s reconstructed Raman signal directly against the ground truth. The network outputs a recovered spectrum that aligns closely with the ground-truth signal. The model resolves the overlapping peaks, rejects the isolated cosmic spike, and produces no spurious peaks in the empty regions of this spectrum. Figure b presents the analysis for the intermediate noise case (SNR = 15). This spectrum incorporates nine distinct peakslocated at 380, 405, 560, 612, 812, 905, 1050, 1300, and 1720 cm^–1which exhibit varying line widths and multiple overlapping regions. As shown in the top panel, the network-generated baseline smoothly tracks the background trend, passing cleanly beneath the prominent Raman features without being skewed by their high intensities. The middle panel compares the baseline-corrected spectrum with the ground truth, illustrating that the relative peak strengths are robustly maintained despite the background removal. Finally, the bottom panel reveals close agreement between the model’s reconstructed Raman signal and the ground-truth data. The network accurately preserves both the amplitude heights and the peak widths, confirming its high fidelity in spectral recovery at moderate noise levels. Finally, the low-noise case (SNR = 25) is demonstrated in Figure c. This spectrum is constructed with 12 Raman peaks located at 360, 405, 560, 612, 735, 805, 901, 1070, 1290, 1480, 1610, and 1730 cm^–1, featuring varying line widths and multiple overlaps. To test the model’s ability to recover signals in extreme edge cases, a synthetic hot pixel was intentionally inserted so that it directly overlaps with the 1480 cm^–1 peak. Consistent with the previous cases, the top panel shows that the generated baseline closely tracks the slowly varying behavior of the raw spectrum, remaining unaffected by the strong peaks and heavily overlapped regions. The middle panel contrasts the baseline-corrected spectrum with the ground truth, where a close match in peak height and width can be observed. The bottom panel demonstrates a close match between the reconstructed Raman spectrum and the ground truth. Notably, the generated Raman signal efficiently filters out the hot pixel present at the 1480 cm^–1 peak while preserving the underlying true Raman feature.

7.

Denoising performance on synthetic data at SNR of (a) 5, (b) 15, and (c) 25. Top row: Input noisy spectra. Middle row: baseline corrected spectrum using the baseline generated by the model overlaid on the ground truth. Bottom row: Comparison between the ground truth and the recovered Raman signal.

Complementing the qualitative visual inspection, Table summarizes the model’s quantitative performance on the three individual spectra of Figure using Mean Squared Error (MSE) and cosine similarity metrics. The metrics in Table and Figure serve distinct purposes and are not directly comparable. Figure illustrates the model’s expected performance through population averages over 500 randomly generated spectra at each signal-to-noise ratio. In contrast, Table provides quantitative annotations for the three individual spectra shown in Figure . These specific spectra were generated independently and contain varying amounts of signal (5, 9, and 12 peaks, respectively). Because mean squared error scales with total signal amplitude, the MSE in Table is influenced by peak content as well as noise. Consequently, it does not decrease monotonically with SNR, unlike the uniform population averages in Figure . The cosine similarity scores indicate a high degree of spectral fidelity, maintaining a value of 0.996 even under the most severe noise conditions (SNR = 5). As expected, this metric steadily improves as noise decreases, reaching 0.9981 at SNR 25. This trend quantitatively confirms the network’s robust capability to preserve the complex morphological structure and relative peak ratios of the underlying Raman signatures. Concurrently, the MSE values remain consistently constrained to the order of 10^–4 across all tested noise tiers. These low error magnitudes indicate that the model not only recovers the correct spectral shape but also accurately reconstructs the absolute intensity values of the Raman signal, maintaining strong amplitude fidelity regardless of the background noise severity.

3. MSE and Cosine Similarity for the Three Individual Spectra of Figure (N = 1 per Row).

SNR	Number of peaks	MSE	Cosine similarity
5	5	2.7 × 10–4	0.996
15	9	5 × 10–4	0.9976
25	12	3.6 × 10–4	0.9981

^aThese are three different spectra containing 5, 9, and 12 peaks respectively. They are not one spectrum measured at three noise levels. Therefore, these values are not comparable with the population averages of Figure . The population averages are computed over 500 randomly generated spectra per SNR.

5.1.1. Contribution of the Individual Components

Each component was removed in turn, with the architecture, the training stream and the frozen validation set otherwise unchanged. The protocol and the per-arm figures are given in Section S4 of the Supporting Information. The run-to-run standard deviation of the reported configuration, measured over three seeds, is 4.5% in mean absolute error, and effects smaller than that are not resolved by these experiments.

The auxiliary deep supervision is the largest contribution among the modules: removing it raises the mean absolute error by 14.2% at an unchanged parameter count. Removing the terminal squelch gate increases the mean absolute error by 7.4%, and withholding the baseline features from the gate increases it by 5.9%. The latent orthogonality penalty and the choice between subtracting and concatenating those features change the error by 3.5% and 1.0%, both within the run-to-run bound, so they are not ranked.

The encoder block shows the largest measured effect. Replacing the RK4 residual blocks with plain preactivation residual blocks raises the mean absolute error by 41.8% and the cosine loss by 88.5%, while reducing the parameter count by 10.2% and the time to process a 96-well plate by 27%. Accuracy and computational cost therefore trade off in this component; the reported configuration retains the RK4 block for accuracy.

5.1.2. Comparison with a Cascaded Architecture

The dual-branch design is defined against the cascaded arrangement. We reproduced the cascaded U-Net of Kazemzadeh et al. and trained it on our own synthetic engine, from the same random seed, on the same stream of training spectra, and evaluated it on the same frozen set of 2 000 validation spectra with the same loss.

The cascade is worse on every recovered-signal metric, by 67.9% in mean absolute error and 162.7% in cosine loss, while carrying 13.9% more parameters (Table and Figure ). The disadvantage is largest in the cosine loss, which measures spectral shape and not amplitude, and largest at low signal-to-noise ratio. In that regime a weak band is most likely to be lost in the first pass and cannot be recovered from the second, because the denoising stage never observes the raw measurement. Section S5.3 of the Supporting Information shows that the amount by which the cascade’s baseline is worse predicts the amount by which its recovered signal is worse, which establishes the mechanism by measurement.

4. Dual-Branch Network against a Cascaded U-Net and a Skip-Free encoder–decoder, Trained and Evaluated on Identical Data.

	Dual-branch	Cascaded	Encoder–decoder
Validation MAE ↓	0.02564	0.04306	0.02588
Validation cosine loss ↓	0.01273	0.03344	0.01234
Baseline MAE ↓	0.00566	0.00635	
Parameters	5 168 630	5 888 130	5 199 041

^aFull results, including behavior across the signal-to-noise range and on experimental spectra, are in Section S5 of the Supporting Information.

8.

Cascaded U-Net and a skip-free encoder–decoder, relative to the dual-branch network on identical data; bars to the right of unity are worse, or larger. The cascade is worse on every recovered-signal metric despite carrying 13.9% more parameters, because detail lost in its first stage cannot be recovered by its second. The encoder–decoder ties the dual-branch network on the recovered signal, but yields no baseline and so supports no independent check.

Section S5 of the Supporting Information states the limits of this comparison. A fully convolutional encoder–decoder trained on the same data matches the dual-branch network on the recovered Raman signal, so the accuracy penalty attaches to the cascaded, two-stage arrangement rather than to depth or parameter count. The encoder–decoder does not emit a baseline; what the dual-branch network uniquely provides is therefore the baseline estimate, the difference-signal check, and I _deep, not a lower reconstruction error than a single-pass denoiser of equal size.

On inference cost the comparison runs the other way. The cascade processes a 96-well plate faster than our network despite performing two forward passes, because the cost of a pass depends on the network executing it and our encoder evaluates its convolutional stack four times per residual block. Every method considered completes a plate in under a second against an acquisition of 10 to 60 s per well, so inference is not what limits screening rate.

5.2. Validation with Real Data

To evaluate transfer from the synthetic engine, the trained network was applied to experimental Raman spectra. The network was trained exclusively on synthetic data, without any exposure to experimental spectra. The trained network recovers experimental spectra, which we attribute to constructing the synthetic engine from empirical background, dark-current and blank spectra acquired on the same instrument. This capability reduces the need for large labeled experimental data sets and, for the samples tested here, for sample-specific transfer learning.

The model was applied to aqueous glycerol at 100, 150, and 200 mM, and to adenine sulfate at 30, 40, and 50 mM in 1 M sodium hydroxide. For each measurement a 330 μL aliquot was dispensed into a 96-well plate and the laser focus optimized at the sample surface. The empirical background library used for training was acquired from water-filled wells, so the adenine samples, and the guanine samples that follow, present the network with a solvent to which it was never exposed. The recovered spectra remain consistent under this change of medium, which is consistent with transfer that is not tied to the training solvent.

Figures a and a show the raw measurement with the network-predicted baseline drawn over it. In both systems the background dominates the acquisition and obscures most Raman features, and in both the predicted baseline follows it smoothly while avoiding the peaks that sit on it. Figures b and b compare the recovered Raman signal with conventional processing, baseline removal by airPLS followed by Savitzky–Golay smoothing. The conventional pipeline fails differently in the two cases: on glycerol at 100 mM it leaves excessive residual noise, which the smoothing filter does not resolve into identifiable peaks. On adenine at 30 mM it returns negative intensities, the signature of an overfitted baseline, and that distortion propagates through the filter. Figures c and c show that in both systems the recovered Raman signal agrees with the difference between the raw spectrum and the predicted baseline, a comparison that does not depend on the Raman branch.

9.

Denoising performance on glycerol samples. (a) Acquired raw spectra for different concentrations 100, 150, and 200 mM compared against the recovered baseline. (b) Comparison between the model Raman signal, the baseline corrected spectrum using airPLS, and the smoothed Raman signal using Savitzky-Golay filter. (c) The recovered Raman signal in comparison with baseline corrected spectrum using our model. The shaded columns represent the reported peaks at 495, 673, 819, 852, 925, 976, 1054, 1115, 1162, 1270, 1466, and 1645 cm^–1. (d) The three dominant peaks tracked for the different concentrations.

10.

Denoising performance on adenine sulfate samples. (a) Acquired raw spectrum for different concentrations 30, 40, and 50 mM compared against the recovered baseline. (b) Comparison between the model Raman signal, the baseline corrected spectrum using airPLS, and the smoothed Raman signal using Savitzky-Golay filter. (c) The recovered Raman signal in comparison with baseline corrected spectrum using our model. The shaded columns represent the reported peaks at 545, 633, 730, 988, 1075, 1135, 1205, 1257, 1334, 1372, 1450, and 1548 cm^–1. (d) The three major peaks tracked for the different concentrations.

The recovered peaks agree with published assignments in both systems. For glycerol, features are detected at 495, 673, 819, 852, 925, 976, 1054, 1115, 1162, 1270, 1466, and 1645 cm^–1. The bands at 976 and 673 cm^–1 correspond to CH₂ rocking and C–C bending, respectively; the envelopes from 800–900 and 1000–1150 cm^–1 lie in the C–O stretching region of the alcohol groups; and the feature at 1645 cm^–1 is the H–O–H bend of the aqueous solvent. Features above 1000 cm^–1 match published assignments for aqueous glycerol. For adenine sulfate they are detected at 545, 633, 730, 988, 1075, 1135, 1205, 1257, 1334, 1372, 1450, and 1548 cm^–1: the dominant band at 730 cm^–1 is the ring-breathing mode of the adenine ring, the band at 1334 cm^–1 the external C–NH₂ stretch, and the band at 988 cm^–1 the symmetric stretch of the sulfate ion, with the remaining features matching previously reported assignments. , Figures d and d track three prominent peaks of each analyte across its concentration series, at 852, 1054, and 1466 cm^–1 for glycerol and at 730, 988, and 1334 cm^–1 for adenine sulfate; both follow a linear trend with concentration, with a slight deviation from linearity at 1466 cm^–1.

To further investigate the model’s performance in quantitative analysis, we prepared seven different concentrations (ranging from 20 to 80 mM) of guanine diluted in 1 M NaOH, each acquired with an integration time of 60 s. Figure a illustrates the model’s performance on a subset of these concentrations (30, 40, and 50 mM). The validity of the peaks detected by the Raman head is confirmed by their alignment with the difference signal (raw spectrum minus recovered baseline). The model successfully recovered major peaks at 654, 1165, 1200, 1230, 1270, 1335, 1380, 1460, and 1540 cm^–1 (highlighted in Figure a). The peak at 654 cm^–1 represents the purine ring breathing mode. Additionally, the peak at 1230 cm^–1 corresponds to the C2-NH₂ stretching mode, while the feature at 1380 cm^–1 represents the C2–N3 stretching mode. The remaining peaks are in agreement with prior studies. , Monitoring I _deep of eq across all seven concentrations reveals a highly linear relationship, achieving an R ² value of 0.9825, as shown in Figure b. Each concentration was prepared as five independent dilutions and measured separately, and the error bars in Figure b are one standard deviation across those five replicates. The relative standard deviation is between 9.5 and 31% from 30 mM upward and rises to 86.8% at the lowest standard, as expected near the detection limit. Quantitative claims are therefore drawn from the range above 30 mM. From the standard deviation of ten blank wells of the solvent alone and the sensitivity of the response near the lowest standard, the limit of detection is 4.2 mM. The replicate measurements are tabulated in the Supporting Information (Section S9). The acquisition time is held constant at 60 s across this series. Thus, the normalization span X _max–X _min is essentially unchanged from sample to sample. This confirms that the observed trend is driven by the reconstructed Raman component itself, rather than the rescaling in eq . To evaluate the model’s stability under varying acquisition times, spectra from a 100 mM guanine sample were acquired using integration times ranging from 10 to 60 s. Figure c compares the Raman signal recovered by our network with the baseline-corrected spectrum. The recovered signal closely tracks the baseline-corrected data, successfully resolving the same major peaks identified during the concentration analysis. Furthermore, the I _deep extracted from the deep Raman layer across the six acquisition intervals demonstrate a robust linear trend, yielding an R ² value of 0.9973. As for the concentration series, each exposure was measured five times independently. The error bars in Figure d represent one standard deviation across those replicates, with the relative standard deviation lying between 4.6 and 12.7%. Unlike the concentration series, both the fluorescence background and the Raman signal accumulate with integration time here. As a result, the normalization span in eq scales with acquisition time. A linear dependence of I _deep on exposure is therefore expected on physical grounds. The reconstructed Raman component remains consistent even as the SNR degrades at the shortest exposures. The readout scales linearly with acquisition time, meaning shorter exposures can be referenced to time-specific calibrations. Because inference is much faster than acquisition (requiring under one second per 96-well plate), any future reduction in acquisition time achieved while preserving a specified analytical performance target would translate directly into improved screening throughput. Detailed precision metrics are provided in Section S9.3.

11.

Denoising performance on guanine samples. (a) The recovered Raman signal compared with the baseline-corrected spectrum for three of the seven concentrations measured between 20 and 80 mM. The shaded columns highlight reported peaks at 654, 1165, 1200, 1230, 1270, 1335, 1380, 1460, and 1540 cm^–1. (b) I _deep (detector intensity units) against concentration. Each point is the mean of five independent dilutions, with error bars indicating one standard deviation. (c) The recovered Raman signal compared with the baseline-corrected spectrum across 6 acquisition times (10 to 60 s). All were recorded on a single 100 mM guanine sample. (d) I _deep against acquisition time for the 100 mM sample. Each point is the mean of five independent measurements, with error bars indicating one standard deviation.

The analytes considered so far are glycerol, a small polyol, and the purine bases adenine sulfate and guanine. To test whether the network transfers to a chemically unrelated class, we measured l-tyrosine, an aromatic amino acid, at four concentrations between 25 and 100 mM in 1 M NaOH with an integration time of 60 s. Neither tyrosine nor any amino acid is represented in the synthetic training distribution, and the network was applied unchanged.

Figure presents the result. The background falls by approximately 1.2 × 10⁴ counts across the window and is not monotonic. Figure a shows the baseline branch following this background without absorbing the superimposed sharp features. Figure b provides a cross-check independent of the Raman branch. Every band in the recovered signal corresponds to a feature in the difference between the raw spectrum and the predicted baseline, and no unsupported bands are introduced. All six characteristic markers of tyrosinate falling within the spectral window are recovered, at 649, 837, 859, 1179, 1214, and 1608 cm^–1. The pair at 837 and 859 cm^–1 is the diagnostic phenol-ring doublet of tyrosine. This doublet is separated by 22 cm^–1 and clearly resolved as two distinct maxima. Figure c shows the recovered signal growing with concentration, and Figure d the integrated signal linear in concentration with an R ² of 0.9993. Signal recovery is therefore consistent across structurally diverse analytes. Recovered features are interpreted against the residual noise σ of the baseline-corrected measurement, using the conventional detection criterion of three standard deviations. , We distinguish features into three practical categories: (1) Assigned, high-confidence analyte features, which exceed 3σ, are present in the difference signal, and exhibit expected concentration dependence. (2) Unassigned reproducible features, which are present but lack a chemical assignment (e.g., the unknown 1073 cm^–1 tyrosine feature). (3) Near-threshold and model-only features, which lie at or below 3σ or appear without raw-data support. Across all four analyte series, every feature reported at or above 3σ was corroborated by the difference signal. Below that level, feature identity is not reliable: in the glycerol series at 100 mM, two bands of comparable strength lie near the threshold and the one reported by the network is not the analyte band. Features below 3σ should therefore be flagged as uncertain, excluded from quantitative interpretation, and checked against the difference signal before being interpreted (see Section S10). Detailed assignments and acquisition conditions are provided in Section S10.1.

12.

l-tyrosine in 1 M NaOH, processed by the unchanged network. (a) The raw measurement at 100 mM with the model-predicted baseline. (b) The recovered Raman signal compared with the difference between the raw spectrum and the predicted baseline. This difference is independent of the Raman branch. (c) The recovered Raman signal at all four concentrations, overlaid. The darkest trace represents the highest concentration. (d) The recovered signal integrated between 400 and 1700 cm^–1 against concentration. This is from a single measurement at each concentration. In (b) and (c), the shaded columns mark the six characteristic markers of tyrosinate that fall within this window, labeled with their recovered positions in cm^–1. The corresponding positions reported for the deprotonated form in aqueous solution are 643, 831, 851, 1174, 1209, and 1601 cm^–1.

6. Conclusion

In conclusion, we have introduced a dual-branch U-Net architecture designed for the simultaneous baseline correction and denoising of Raman spectra. By using a shared encoder that diverges at the bottleneck into dedicated baseline and Raman recovery decoding heads, our model effectively avoids distortions that arise from sequentially applying denoising and baseline correction techniques. A distinguishing feature of this architecture is the integration of a cross-branch gating mechanism, which provides spatial peak localization hints by evaluating feature differences between the encoder and the baseline decoder.

The model was trained using a custom-developed synthetic Raman engine configured to simulate spectra acquired by the RamanBot microplate platform. Comprehensive validation on these synthetic data sets demonstrated robust peak recovery across a broad range of SNRs, preserving high spectral fidelity even under severe noise conditions.

The model was validated on experimental data. The network retrieved the Raman signal from highly noisy glycerol samples and from moderately noisy adenine sulfate samples, and from l-tyrosine as an independent analyte class. Furthermore, our results confirmed that the network preserves relative peak intensities, enabling reliable quantitative analysis. We benchmarked our dual-branch approach against the conventional airPLS baseline correction and Savitzky–Golay smoothing pipeline. Our model recovered the Raman signal with less residual noise and avoided the baseline overfitting observed for airPLS on the adenine sulfate samples. Benchmarked against a cascaded U-Net trained on the same data, the dual-branch network was more accurate on every recovered-signal metric: the cascade’s validation mean absolute error was higher by 67.9% and its cosine loss by 162.7%, while it carried 13.9% more parameters. Finally, integrating I _deep at the deep Raman layer yielded a linear response (R ² of 0.9825 for concentration and 0.9973 for acquisition time, across five replicates), with a limit of detection of 4.2 mM. These results indicate that the model improves the effective SNR. I _deep scales linearly with acquisition time between 10 and 60 s. A calibration built at a given exposure therefore supports screening at that exposure; a 10 s integral cannot be rescaled to a 60 s integral by the ratio of the exposures. Because acquisition, not inference, dominates the plate time budget, screening at 10 s rather than 60 s increases the number of wells measured per unit time by a factor of 6, provided a calibration is built at that exposure.

The code, the trained weights of the reported model, and the measured spectra supporting this study are available at https://github.com/khaledAtia/Raman_Denoising. The release includes the model definition and training script, the synthetic data engine and the empirical background and noise banks it draws on, the evaluation and figure-generation scripts, and the raw experimental spectra used in this study. Its contents are itemized in Section S12 of the Supporting Information.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.6c04490.

• The Supporting Information is available free of charge on the ACS Publications Web site. Architecture details and a formal definition of the cross-branch gating module, with a layer-by-layer specification of the network (Section S1); full parameterization of the synthetic data engine (Section S2); training protocol, data splits, and what the training and validation distributions share (Section S3); ablation study covering the auxiliary deep supervision, the terminal squelch gate, the gate input, the latent orthogonality penalty, the encoder residual block, the bottleneck channel allocation, the encoder width, and the loss weights (Section S4); comparison against a cascaded U-Net and a skip-free encoder–decoder on synthetic and experimental spectra, and across the signal-to-noise range (Section S5); inference timing on a processor and on a graphics processor, against the classical pipeline (Section S6); deep-layer spatial resolution, the width dependence of the recovered integral, and its mass balance between the two decoder branches (Section S7); definition and calibration of the deep-layer intensity integral I _deep (Section S8); quantitative validation with replicates, precision, and detection limits (Section S9); additional experimental validation, including behavior near the detection limit in the glycerol series (Section S10); the relationship between the population statistics of Figure and Table (Section S11); data and code availability (Section S12); and an index of the supporting figures (Section S13) (PDF)

The authors declare no competing financial interest.