Phase-aware adaptation for functional group recognition from condensed-phase infrared spectra

Abstract

Infrared (IR) spectroscopy is widely used for functional group identification, but machine-learning models often struggle when spectra are transferred across physical phases. Here we establish a phase-separated 16-label benchmark for functional group recognition across gas, liquid, and solid IR spectra and show that direct gas-to-condensed transfer is unreliable, especially for solid spectra. Simply increasing model capacity or replacing the backbone provides only limited improvement. In contrast, the main two-stage gas-pretrained adaptation protocol substantially improves condensed-phase recognition over target-only training, raising solid Macro-F1 from 0.6902 to 0.7351 and liquid Macro-F1 from 0.7198 to 0.7792. The improvement is class-dependent, with strong recovery for labels such as nitrile, while alkene remains difficult because of weak and overlapping spectral evidence. Additional analyses show that the adaptation benefit is not explained by longer target-only training or class weighting, and that it transfers across multiple model families. To make the adapted model more suitable for resource-limited use, we applied segment max pooling under the two-stage gas-pretrained adaptation pipeline, reducing the model to 9.23% of the original Dual Head parameter count while retaining 99.04% of its Macro-F1 performance. Together, this work establishes a phase-aware adaptation framework for condensed-phase IR functional group recognition, showing that gas-phase spectra can be converted from an unreliable direct predictor into useful prior training data for solid and liquid spectra.

Gas-phase spectra become useful prior knowledge when adapted to condensed-phase environments, improving functional-group recognition across solid and liquid IR spectra despite phase-dependent spectral variation.

1. Introduction

Infrared (IR) spectroscopy is one of the most widely used techniques for molecular characterization.^1–4 Molecular vibrations are sensitive to bond types, local chemical environments, and functional group identity, so IR spectra carry useful information about molecular structure. In routine organic, materials, and analytical chemistry, IR spectroscopy provides a rapid means of identifying common functional groups and supporting structural assignments. IR alone is usually not enough for full structure elucidation, but functional group recognition from IR spectra remains an important and practical part of spectral interpretation.

Traditional IR interpretation relies heavily on expert knowledge of characteristic absorption regions and peak patterns. A broad absorption may point to hydrogen-bonded hydroxyl groups, a strong band in the carbonyl region may indicate a carbonyl-containing group, and diagnostic peaks in the fingerprint region can add further structural evidence.^5,6 Manual interpretation becomes harder when spectra contain overlapping bands, weak diagnostic features, baseline variation, or several functional groups with related absorption patterns. These difficulties make automated IR interpretation an attractive task for machine learning. A model that identifies functional groups directly from spectral input could give a useful first-pass annotation and make spectral analysis more consistent.^7,8

Recent studies have applied machine-learning and deep-learning methods to several IR analysis tasks, including functional-group identification,^7,9–13 automated materials and molecular-structure characterization,^8,14,15 and mixture or composition analysis.^16–18 Both convolutional neural networks^8,10,11 and transformer-based architectures^9,12,18 have achieved strong performance in functional-group recognition and related IR classification tasks, demonstrating the potential of data-driven methods to support automated IR analysis.

An important limitation of current automated IR functional-group recognition studies is that physical phase is rarely treated as an explicit evaluation variable.^9–11,14 Some models are developed and tested only on gas-phase spectra, while others use datasets for which phase information is unavailable or not reported in sufficient detail. As a result, strong benchmark performance does not necessarily establish whether a model remains reliable on the liquid- and solid-phase spectra, which are more commonly encountered in laboratory practice.¹⁹ For gas-trained models, it is unclear whether the learned spectral patterns can be transferred directly to condensed phases. For phase-unresolved datasets, overall performance cannot reveal how recognition accuracy differs among gas, liquid, and solid spectra or whether failures are concentrated in a particular phase.

The spectral consequences of physical phase are substantial. In the gas phase, relatively free molecular rotation can superimpose rotational structure on vibrational absorptions, so individual vibrational bands may contain resolved or partially resolved P- and R-branch features rather than appearing as a single smooth band.²⁰ Whether this rotational structure is resolved depends on factors including molecular rotational constants, molecular symmetry, spectral congestion, and instrumental resolution. In liquids, frequent collisions and intermolecular constraints hinder free rotational motion,²¹ so the discrete gas-phase P- and R-branch structure is generally not resolved in liquid-phase bands. In solids, molecular rotation is strongly restricted by the crystal lattice. Across both condensed phases, intermolecular interactions,²² hydrogen bonding and solvation,²³ vibrational coupling,²⁴ and crystal packing²⁵ can shift vibrational frequencies, alter band widths and shapes, and redistribute relative intensities. Characteristic frequency regions associated with a functional group may therefore remain informative across phases, while detailed band shapes, widths, exact frequencies, and relative intensities can differ substantially. A model trained on one phase may learn associations that do not remain reliable under another spectral environment.

These considerations raise two unresolved questions. First, can models trained on gas-phase IR spectra be applied directly to liquid- and solid-phase spectra? Second, if direct transfer is unreliable, can condensed-phase recognition be improved by introducing gas-phase spectra as prior training data into existing models, without requiring the development of a new model architecture? Answering these questions requires phase-separated evaluation, because the usefulness of gas-derived information cannot be distinguished from phase-specific performance when the phases are combined or left unspecified.

In this work, we use the previously developed Dual Head CNN model¹⁰ as the primary experimental framework to establish a phase-separated benchmark for multilabel functional group recognition across gas-, liquid-, and solid-phase IR spectra under a consistent 16-label definition. Fig. 1 provides an overview of the phase-aware framework, from the failure of direct gas-to-condensed transfer to two-stage gas-pretrained adaptation, 16-label functional-group prediction, and parameter-efficient model compression. We first evaluate whether a model trained on gas-phase spectra can be applied directly to liquid and solid spectra. We then test whether the limitations of condensed-phase recognition can be overcome through model scaling or backbone replacement.

Fig. 1. Overview of the phase-aware adaptation framework for condensed-phase infrared functional group recognition. A model trained only on gas-phase spectra performs well within the gas phase but transfers unreliably to solid- and liquid-phase spectra. In the two-stage gas-pretrained adaptation protocol, the model is first trained jointly on gas and target-phase spectra and is then fine-tuned using target-phase spectra alone. The adapted model performs multilabel prediction across 16 functional-group labels. Segment max pooling can further compress the spectral representation and substantially reduce the model parameter count while preserving most of the adapted-model performance.

We next investigate whether gas-phase spectra can instead be used as prior training data to improve existing models without introducing a new architecture. In the proposed adaptation strategy, gas spectra provide an initial source of functional-group information, after which the model is adapted to the target condensed phase. This strategy is compared with target-only training on both solid and liquid spectra. Class-level evaluation and spectral analyses are used to examine where the gas prior succeeds and where it remains limited. We further test whether the benefit extends to other convolutional and attention-based architectures and whether the adapted model can be compressed while preserving its performance. In this work, joint gas–target training followed by target-phase fine-tuning is termed two-stage gas-pretrained adaptation, or simply two-stage adaptation. When an additional low-rate target-phase refinement stage is introduced, the resulting procedure is termed three-stage gas-pretrained adaptation, or simply three-stage adaptation.

Direct gas-to-condensed transfer proved unreliable, with the largest degradation observed on solid spectra, while increasing model capacity or changing the backbone provided only limited recovery. By contrast, two-stage gas-pretrained adaptation increased Macro-F1 from 0.6902 to 0.7351 on solid spectra and from 0.7198 to 0.7792 on liquid spectra. The improvement was class-dependent, with strong recovery for nitrile but persistent difficulty for alkene. The benefit was also observed across several convolutional and attention-based models, showing that it was not restricted to a newly designed architecture. Segment max pooling further reduced the adapted Dual Head model to 9.23% of its original parameter count while retaining 99.04% of its Macro-F1 performance. These results show that gas-phase spectra are unreliable as direct predictors of condensed-phase spectra but can serve as useful prior training data when physical phase is treated explicitly.

2. Results and discussion

2.1. The gas-to-condensed phase gap

IR spectra were retrieved from the NIST Chemistry WebBook and separated by the sampling state recorded in the original records, giving 5743 gas-phase, 1484 solid-phase, and 420 liquid-phase spectra.²⁶ All spectra were resampled onto a common wavenumber interval of 3844 to 552 cm⁻¹ at 2 cm⁻¹ spacing and scaled to the range 0 to 1, following the procedure of our previous work.¹⁰ Functional groups were assigned independently for each spectrum from its associated molecular structure using RDKit and the same SMARTS definitions as in previous works. Gas, liquid, and solid records were therefore labeled directly from their own associated structures. Sixteen labels with sufficient representation in the condensed-phase subsets were retained for the present multilabel benchmark. All models were evaluated by five-fold cross-validation on the target phase, with the held-out fold used only for testing. Full details of the dataset and the training protocol are given in the SI. Models following the common training protocol were optimized using a combination of binary cross-entropy and soft-F1 loss, as detailed in the SI.

A natural first question is what happens when a model trained on gas-phase spectra is applied directly to solid and liquid spectra, without any further training on the target phase. On its own task the model is strong. It reaches a five-fold Macro-F1 of 0.9089 with a standard deviation of 0.0054, and a Micro-F1 of 0.9278 with a standard deviation of 0.0030 on gas spectra. Gas-phase functional group recognition is therefore learned well, and any later degradation reflects the change of phase rather than a weakness of the model. The gas-phase reference values were obtained by five-fold cross-validation, whereas the direct-transfer results below were obtained from a single model trained on the full gas-phase dataset and evaluated once on the complete solid- and liquid-phase subsets.

The degradation is large. On solid spectra, Macro-F1 falls to 0.4281 and Micro-F1 to 0.5878. On liquid spectra the same model reaches 0.5933 and 0.7395. The two condensed phases are not affected equally. Solid recognition collapses further than liquid, and this ordering recurs throughout the per-class results.

The per-class results show that the loss is uneven rather than a uniform downward shift. A few classes survive the phase change. Alkane and aromatic stay recognizable in both condensed phases, and on liquid spectra a broader set also holds up moderately well, including amine, nitro, and the aromatic-substitution labels. Other classes give way almost entirely. On solid spectra, alkene, organohalide, ketone, carboxylic acid, and amide all fall below an F1 of 0.20, even though each scored above 0.81 on gas. Liquid spectra recover some of the solid losses, and the classes that remain weakest are largely the same.

Fig. 2 summarizes this class- and phase-dependent pattern. The pattern, not just the averages, is what matters here. The failures are class- and phase-dependent rather than uniform. A gas-trained model applied directly to condensed-phase spectra therefore cannot be trusted as a functional group assigner, because which groups it identifies correctly changes from one phase to another, and the groups it loses are often chemically central ones. Whether gas-phase information can instead be made useful through phase adaptation, rather than direct deployment, is the question the rest of this work takes up.

Fig. 2. Per-class F1 scores of the gas-trained Dual Head model evaluated on gas-, liquid-, and solid-phase IR spectra. Performance remains consistently high on gas-phase spectra but decreases substantially and unevenly under direct transfer to both condensed phases. Solid-phase performance is lower overall than liquid-phase performance, although the magnitude of the solid–liquid gap varies across classes. Particularly low condensed-phase F1 scores are observed for alkene, organohalide, ketone, carboxylic acid, and amide. These results demonstrate that strong gas-phase recognition does not translate reliably to direct condensed-phase inference.

2.2. Larger models and different architectures give limited gains

Direct gas phase inference is unreliable on condensed phase spectra, so we next asked whether the problem could be addressed at the model level, by increasing model capacity or by replacing the architecture. Two groups of models were compared. The first group consisted of models trained in this work, covering the Dual Head family, a ResNet²⁷ baseline, and a patch transformer of our own design using a Transformer encoder.²⁸ The second group consisted of the published architectures AggMapNet¹¹ and IRCNN.⁸ All models were evaluated on the same phase specific data splits. The Dual Head variants shared a common preprocessing workflow and training protocol. The ResNet and patch transformer models used comparable training procedures adapted to each architecture, so that capacity and architectural changes could still be compared under a consistent evaluation protocol. The published architectures were instead trained following the strategy described in their original work. Full implementation details, including the preprocessing workflow, the training protocol applied to each model type, and the architectural specifications, are provided in the SI.

Increasing capacity within the Dual Head family produced only modest changes. The baseline Dual Head model reached a Macro-F1 of 0.6902 on solid spectra and 0.7198 on liquid spectra. Widening the MLP head raised solid Macro-F1 to 0.6963 but left liquid performance flat at 0.7187. Doubling the channel width improved solid Macro-F1 to 0.7044 and liquid Macro-F1 to 0.7322. A wider MLP head combined with doubled channels gave the best solid result in this group at 0.7084, yet the gain over the baseline was only 0.0182 even though the parameter count rose from 27.94 million to 111.74 million. Adding an extra final convolution did not help, giving 0.6931 on solid and 0.7164 on liquid. Larger Dual Head models can therefore offer small improvements, but the benefit saturates quickly and does not scale with parameter count.

Replacing the architecture did not provide a robust solution either. Among the published architectures, AggMapNet reached a solid Macro-F1 of 0.6457 and a liquid Macro-F1 of 0.6792, and IRCNN reached 0.6125 and 0.6232 on the two phases. Both stayed below the Dual Head family despite the large parameter count of IRCNN.

The patch transformer we designed reached a solid Macro-F1 of 0.5277 and a liquid Macro-F1 of 0.5414. Its per class behavior differed from that of the ResNet baselines. It left only one label at zero on solid spectra and no label at zero on liquid spectra, whereas the ResNet variants dropped several labels to zero or close to zero on both phases. This is likely because its convolutional patch embedding and pooled readout suit the local peak structure of IR spectra. Even so, it stayed below the Dual Head family, which shows that a task appropriate design narrows the gap but does not close it.

The ResNet based models failed regardless of size. Under the same target-only protocol, the larger variant reached Macro-F1 values of 0.3488 on solid spectra and 0.2257 on liquid spectra, while the smaller variant reached 0.2961 and 0.2400. Reducing the parameter count did not recover performance, and the Dual Head family remained stable even at 111.74 million parameters, so the collapse cannot be attributed to capacity alone. The training and validation losses of both ResNet models were monitored across training (Fig. S2). The training loss falls steadily while the validation loss stops improving early, so the low scores do not reflect insufficient training. The behavior instead points to a mismatch between the ResNet-50-1D architecture and this task, where a deep residual backbone designed for large scale settings does not suit the small and noisy condensed phase data.

As shown in Fig. 3, model size alone did not explain performance. AggMapNet used only 0.43 million parameters yet outperformed IRCNN at 63.98 million, and the patch transformer used 0.89 million parameters yet outperformed both ResNet variants by a wide margin. The ResNet variants were the weakest overall regardless of their size. Performance did not track parameter count in either direction, which points to architecture and training data rather than capacity as the more important factors. The condensed phase data are limited in number, higher in noise, and long tailed in label distribution, and these properties appear to matter more than model size. Difficult labels such as alkene and nitrile did not recover with larger capacity, and nitrile in particular fell to zero or close to zero in several models. The limitation is therefore not a shortage of parameters. Condensed phase IR recognition instead appears to need a training strategy or a phase aware prior that can compensate for scarce and noisy data rather than a larger or different backbone.

Fig. 3. Comparison of model capacity, architecture, and condensed-phase recognition performance across the evaluated models. The upper panel shows Macro-F1 scores on solid- and liquid-phase spectra together with the corresponding parameter counts on a logarithmic secondary axis. Models are grouped into the Dual Head family, published architectures, ResNet-50-1D variants, and the patch transformer. Increasing the size of the Dual Head model produces only modest gains, while parameter count does not show a consistent relationship with performance across architectures. The lower panel shows per-class F1 scores for alkene, ester, and nitrile on solid and liquid spectra. Phase is distinguished by line style, while class identity is distinguished by color and marker. Difficult-class performance varies substantially across architectures, with several models showing near-zero F1 for nitrile or pronounced degradation for alkene and ester. Overall, neither increasing model capacity nor replacing the backbone provides a general solution to condensed-phase functional group recognition.

The per class results within the Dual Head family confirm this reading. Larger capacity did not produce a consistent recovery on the difficult labels. Alkene stayed near the middle of the 0.5 range on both solid and liquid spectra across every variant, and nitrile on solid spectra fluctuated widely from one model to the next without a clear link to parameter count. The labels that limit condensed phase performance were therefore not resolved by adding capacity within a fixed architecture.

These results suggest that modifying an existing framework or designing a new architecture may not be the most effective route to the condensed phase problem. Training on the target phase alone gives only small, class dependent, and bounded gains from added capacity. They point instead toward a different direction, where prior knowledge from gas phase spectra might improve recognition on the smaller, noisier, and more complex condensed phase data at a lower development cost. We examine this possibility in the next step, where gas phase spectral information is adapted into the target phase rather than scaling a model trained on the target phase alone.

2.3. Two-stage gas-pretrained adaptation improves target-phase recognition

The results above suggest that the condensed-phase limitation is not mainly solved by increasing model size or replacing the backbone. We therefore tested whether gas-phase spectra could serve as prior training data that is then adapted to each target phase. For each target phase, the target dataset was first divided into five folds. In each fold, the held-out target split was used only for final evaluation. The model was first trained on gas spectra together with the in-fold target training spectra and was then further refined on the same target-phase training spectra alone. The mixed gas–target training stage and the subsequent target-only fine-tuning stage were each run for 15 epochs. Gas spectra therefore acted not as a directly deployable predictor, but as a source of spectral prior information that was reshaped by target-phase training.

This strategy improved recognition on both condensed phases relative to direct target-phase training. On solid spectra, the 15-epoch solid-only Dual Head baseline reached a Macro-F1 of 0.6902 with a standard deviation of 0.0416 and a Micro-F1 of 0.7749 with a standard deviation of 0.0140. After two-stage adaptation, the same backbone reached a Macro-F1 of 0.7351 with a standard deviation of 0.0170 and a Micro-F1 of 0.7919 with a standard deviation of 0.0100. On liquid spectra, the target-only baseline reached a Macro-F1 of 0.7198 with a standard deviation of 0.0403 and a Micro-F1 of 0.8107 with a standard deviation of 0.0217. Two-stage adaptation raised these values to 0.7792 with a standard deviation of 0.0420 and 0.8579 with a standard deviation of 0.0238. The overall improvements were significant on both phases under fold-level paired tests. Direct gas-to-condensed inference was unreliable, yet the same gas spectra became useful once incorporated through target-phase adaptation.

Using the same fold-wise pairing, we further quantified the uncertainty of the representative class-specific changes. On solid spectra, the mean paired F1 changes were +0.2597 for nitrile (95% CI, −0.0532 to 0.5726) and +0.0513 for alkene (95% CI, −0.0048 to 0.1074). On liquid spectra, the corresponding changes were +0.2165 for nitrile (95% CI, −0.0307 to 0.4636) and +0.1423 for alkene (95% CI, 0.0600 to 0.2247). The wider confidence intervals for nitrile reflect substantial fold-to-fold variation, consistent with the limited numbers of nitrile-positive spectra in the solid and liquid subsets (32 and 21, respectively). Complete fold-level paired results for the overall metrics and all 16 functional-group labels are provided in Section S9 of the SI.

The per-class results show that this improvement was not a uniform shift across all labels (Fig. 4). For solid spectra, adaptation preserved the already strong performance of the alkane and aromatic labels and improved several weaker classes. Organohalide rose from 0.4567 to 0.5196 in F1, carboxylic acid from 0.6690 to 0.7449, nitro from 0.7540 to 0.8044, amide from 0.6190 to 0.6743, and nitrile from 0.5886 to 0.8483. The nitrile gain was the largest in the set. Not every class moved in the same direction. Ether changed little and ketone decreased slightly. Alkene rose from 0.5625 to 0.6138 but stayed among the lowest-scoring labels, a pattern examined in the next section.

Fig. 4. Per-class comparison between target-only training and two-stage gas-pretrained adaptation on condensed-phase IR spectra. (A) F1 scores on solid-phase spectra. (B) F1 scores on liquid-phase spectra. Two-stage gas-pretrained adaptation improves Macro-F1 on both phases, but the gains are class dependent rather than uniform. The largest improvements are observed for nitrile in both phases, together with clear gains for organohalide, carboxylic acid, nitro, amide, and alkene on solid spectra and for aromatic, ester, and alkene on liquid spectra. Several classes change only slightly or decrease modestly after adaptation, showing that the gas prior selectively reorganizes class-level recognition rather than producing a uniform upward shift across all labels.

Liquid spectra showed a similar overall trend with a different class-level pattern. The clearest gains were nitrile from 0.7502 to 0.9667, aromatic from 0.8683 to 0.9339, ester from 0.7950 to 0.8925, and para from 0.7686 to 0.8417, with smaller increases for ortho, meta, and amine. Ketone, amide, and nitro were unchanged or slightly lower after adaptation. Alkene again rose, from 0.5378 to 0.6801, while remaining one of the weakest labels.

Taken together, the gains were concentrated in selected labels rather than distributed uniformly across the label set. Several of the weakest classes remained difficult after adaptation, motivating a closer examination of the underlying success and failure modes.

Beyond aggregate and per-class scores, we examined how two-stage adaptation reshaped the model output in the decision space. For each target phase, paired class logits were plotted for samples grouped by their true-label combinations, with the target-only model as the pre-adaptation condition and the adapted model as the post-adaptation condition (Fig. 5).

Fig. 5. Paired class-logit distributions before and after two-stage gas-pretrained adaptation for solid- and liquid-phase spectra. The upper panels compare alkene and aromatic logits, while the lower panels compare ether and nitrile logits. Samples are grouped according to their true label combinations, and the dashed lines indicate the zero-logit decision thresholds. After adaptation, part of the alkene–aromatic co-labelled population shifts toward positive alkene logits, although a subset remains misclassified. Nitrile-positive samples show a clearer movement toward strongly positive nitrile logits, consistent with the large per-class F1 improvement, whereas the change in ether logits is less clearly reflected in its final F1 score.

The first pair compares alkene and aromatic, two labels that both carry C C-related spectral information. Samples carrying only one of the two labels separated along opposite axes, with aromatic-only samples at positive aromatic logit and alkene-only samples at positive alkene logit. Many samples carrying both labels showed a positive aromatic logit but a negative alkene logit, so aromatic was recognized while the accompanying alkene group was missed. After adaptation, part of this co-labelled population shifted toward positive alkene logits, although a subset remained strongly negative. The same pattern appeared on liquid spectra. The residual alkene difficulty is therefore concentrated in spectra where alkene evidence coexists with aromatic signatures.

The second pair compares ether and nitrile, two labels associated with comparatively distinct diagnostic regions, so each class is read on its own axis rather than as a confusion pair. After adaptation, nitrile-only samples shifted toward strongly positive nitrile logits, consistent with the large increase in nitrile F1 reported above. Ether-only samples also shifted toward more positive ether logits, yet the solid-phase ether F1 changed little. The clearer separation in the decision space was not always accompanied by a corresponding increase in per-class F1.

Across both class pairs, adaptation increased the magnitude of many logits. This sharpened confident correct predictions, particularly for nitrile, but it also increased confidence in some residual errors, which left the affected samples far from the decision threshold. Two-stage adaptation therefore did not shift all outputs toward the positive class. It reorganized the class-specific decision structure instead.

The performance and logit results show that two-stage adaptation improves recognition and reshapes the class-level decision structure, but they do not show where inside the model this change occurs. An improvement in prediction can result either from adjustment of the final classifier or from changes in the features learned throughout the network, and these two routes carry different meanings for what the gas prior contributes. To separate them, we compared the adapted model against the target-only model block by block, using three quantities that each capture a different aspect of the change (Fig. 6). Within each fold the two models started from the same initialization and used the same target-phase training split, so the gas prior is the only difference between them. All values are averaged over the five folds and are reported for both condensed phases.

Fig. 6. Block-wise comparison between target-only and two-stage gas-pretrained adapted models on solid- and liquid-phase spectra. Relative activation drift, final relative parameter distance, and update direction cosine are reported for each model block. In both phases, the earliest convolutional blocks change least, whereas later feature-processing blocks and the classifier show larger divergence. The classifier head has the largest parameter distance and the lowest update-direction agreement, indicating that adaptation follows a distinct optimization path rather than simply extending target-only training.

The first quantity is the relative parameter distance between the two final models within each block, taken relative to the magnitude of their shared initialization. It measures how far apart the two trained models ended up in each part of the network. On both condensed phases the parameter distance was largest in the classifier head, reaching 1.2587 on solid spectra and 1.0720 on liquid spectra, an order of magnitude above every other block. Within each convolutional branch the distance grew with depth, and the first block of each branch barely moved, reaching 0.0232 for Local block 1 and 0.0289 for Global block 1 on solid spectra, with 0.0234 and 0.0331 on liquid spectra. The refinement block and the last global block sat together at the top of the convolutional range, at 0.1608 and 0.1607 on solid spectra. Parameter distance alone, however, only reports how far the two models are from each other. It does not indicate whether the two training protocols were pursuing the same solution or diverging toward different ones.

The second quantity, the update direction cosine, addresses that question directly. It measures whether target-only training and the two-stage adaptation protocol move the parameters in the same direction from their shared initialization. A value close to 1 means the two protocols make similar updates, and a value near 0 means the two updates are nearly orthogonal and follow different optimization paths. The cosine stayed positive in every block on both phases, so the two-stage updates did not reverse the target-only direction. The values were nonetheless modest across the network, remaining at or below 0.5669 on solid spectra and 0.4520 on liquid spectra, which places the two-stage updates on a different path from the target-only updates rather than along the same direction. The classifier head gave the lowest cosine on both phases, 0.0899 on solid spectra and 0.1240 on liquid spectra, while also giving the largest parameter distance. The two routes separate most in the classifier head and agree least on the direction taken to get there.

The two parameter-space quantities describe how the weights moved, but not whether these movements changed what the blocks actually compute. The third quantity, the activation drift, closes this gap by measuring how much the block output features on the same test spectra differ between the two models. On both phases the drift was small at the first block of each branch, 0.2046 and 0.2220 on solid spectra, and rose along both branches to 0.9001 and 0.8529 at the third block. Parameter distance and activation drift do not track each other. Local block 3 moved only 0.0963 in parameter space on solid spectra, which places it in the lower half of the convolutional blocks, yet its output features drifted by 0.9001, above every global block. A small weight change in this block therefore still altered what the block computes. The two phases differ at the end of the network. On solid spectra the drift peaked at the refinement output, at 1.0760, and fell to 0.9123 at the logits. On liquid spectra the drift rose instead to 1.4309 at the logits, the largest value recorded on either phase. Adaptation reshapes the fused spectral representation on solid spectra and the class decision layer on liquid spectra.

The three quantities show that two-stage adaptation changes more than the final classifier. On both phases the first block of each convolutional branch stays close to the target-only model in weights and in output features, whereas the later convolutional blocks, the fused representation, and the classifier head carry the largest changes in what they compute. The effect of including gas-phase data is therefore unevenly distributed with depth rather than confined to the final classifier.

2.4. Evaluation and visualization of class-specific success and failure modes

The per-class results in the previous section show that the gas prior does not act uniformly. Nitrile recovered strongly while alkene stayed among the weakest labels. Splitting the per-class F1 into precision and recall makes the difference explicit. On solid spectra, nitrile precision remained high, decreasing only slightly from 0.9667 before adaptation to 0.9500 after, while nitrile recall increased from 0.4821 to 0.7885. Thus, adaptation recovered many missed nitrile positives with only a small loss of precision. Alkene behaved differently. Alkene recall rose from 0.4760 to 0.5444, a smaller gain, while alkene precision fell from 0.8032 to 0.7069. Across the sixteen labels, alkene showed the largest precision loss among the classes whose F1 improved. The same protocol therefore improved nitrile recall with little loss of precision, whereas the alkene recall gain was accompanied by a more substantial precision decrease.

We applied an occlusion analysis to a nitrile-containing solid spectrum that the adapted model classified correctly. A narrow window was slid along the wavenumber axis, the absorbance inside the window was set to zero, and the masked spectrum was passed through the model, with the change in the target-label logit recorded at each window position. Full window and smoothing settings are given in the SI. Masking the C N stretching region near 2250 cm⁻¹ reduced the nitrile logit from 19.72 to below zero and flipped the prediction to negative. Masking the carbonyl control region between 1650 and 1800 cm⁻¹ left the logit almost unchanged, and masking the alkane C–H region did not reduce it. For this spectrum the correct call rested on the C N band and not on other regions. The high precision retained after adaptation indicates that the recall gain was not accompanied by a substantial increase in false-positive predictions.

Alkene has a different spectral situation. Its C C stretch is weak, and its C–H stretch above 3000 cm⁻¹ lies next to the saturated C–H region, so both features sit close to absorptions that other labels also carry.⁵ Solid samples carrying both alkene and aromatic groups concentrated at high aromatic logit and negative alkene logit before adaptation, and a subset of them remained at strongly negative alkene logit afterwards. Alkene recall was below alkene precision in every setting examined here, including gas spectra in previous work,¹⁰ where the same missed-detection pattern was attributed to overlap between C C and C–H stretching modes and other hydrocarbon functionalities. The same imbalance is already present before any phase change, and adaptation reduces it in the condensed phases without removing it. With the diagnostic features shared in this way, the positives could not be moved without the negatives moving as well, which is what the loss in precision records.

The representative nitrile occlusion result is shown in Fig. 7. Occlusion on individual alkene spectra shows how this appears in practice. Fig. 8 shows two alkene samples that the adapted model did not recover. In Panel A the alkene C–H stretch above 3000 cm⁻¹ is difficult to distinguish from the neighboring saturated C–H region, so this part of the spectrum offers no clear position for the model to anchor on. The diagnostic information that remains lies in the fingerprint region, where the peaks are dense and heavily overlapped, and the strongest logit changes do not fall there. The model instead responds most to the broad absorption near 3300 cm⁻¹ and to the absorption near 2900 cm⁻¹, neither of which is specific to alkene. In Panel B the absorption near 3000 cm⁻¹ appears as two broad peaks merged together, and the alkene C–H stretch is not resolved from them. The strongest logit changes fall near 2000 cm⁻¹, a region that carries almost no absorption. In both spectra, the regions the model relied on do not fall on alkene-specific positions. This response is unlikely to represent a diagnostic alkene absorption and instead suggests that the model is sensitive to non-specific or distributed spectral features in this example.

Fig. 7. Representative solid spectrum colored by the scaled change in nitrile logit after two-stage gas-pretrained adaptation. The strongest change is localized near the characteristic C N absorption region around 2220–2260 cm⁻¹, while most other spectral regions show smaller effects.

Fig. 8. Occlusion analysis of two alkene-containing solid-phase spectra that remained misclassified after two-stage gas-pretrained adaptation. (A and B) Absorbance spectra are colored according to the scaled change in alkene logit caused by masking each spectral region, with the corresponding molecular structures shown as insets. In both examples, the regions producing the largest logit changes do not align consistently with clearly resolved alkene-specific absorptions, illustrating the weak and overlapping spectral evidence available for alkene recognition.

Alkene recognition on this benchmark therefore appears to be limited by the spectral evidence rather than by the transfer strategy. A prior learned from gas spectra can restore predictions that the target-phase model failed to make, but it cannot supply a diagnostic feature that the spectrum does not separate. Alkene remains an open problem on this benchmark, and closing it is likely to require information beyond a single IR spectrum such as complementary NMR data.

2.5. Dissecting late-stage refinement and phase-dependent behavior

To extend the two-stage adaptation into a three-stage protocol, we appended a short low-rate target-phase refinement stage. Five consistently weak classes, namely alkene, amide, ether, ester, and carboxylic acid, were up-weighted in the initial refinement variant. This refinement raised the solid Macro-F1 from 0.7351 to 0.7458 and the liquid Macro-F1 from 0.7792 to 0.7987. To test if the gain is credited to the weighting, we repeated the refinement with equal class weights and held everything else fixed. The unweighted variant reached a solid Macro-F1 of 0.7455 and a liquid Macro-F1 of 0.7991. It matched the weighted variant to within a thousandth of a Macro-F1 point. The two variants differed in opposite directions across the phases.

The class weighting also failed to move its intended targets. Among the five up-weighted classes, weighting shifted the solid per-class F1 by at most 0.012 in either direction. It raised alkene and ether, lowered amide and ester, and left carboxylic acid unchanged. The liquid targets behaved no more consistently. Ester rose, alkene and ether fell, and amide and carboxylic acid stayed unchanged. The weighting produced no coherent effect even on the classes it was designed to favor. The improvement over the two-stage adaptation comes from the low-rate refinement, not from the weighting.

The refinement results suggest that additional target-phase training can contribute to the final performance. We therefore asked whether extending target-phase training alone could reproduce the benefit of two-stage adaptation.

Pretraining epochs and target-only training epochs are not directly comparable, because they optimize on different data distributions and serve different roles. The 30- and 35-epoch target-only controls were therefore not intended to match the total epoch count of the two-stage gas-pretrained protocol. Instead, they were used to test whether simply training the target-only baseline for longer could remove the need for a gas-derived prior.

On solid spectra, the target-only model increased from a Macro-F1 of 0.6902 at 15 epochs to 0.7060 at 30 epochs but decreased again to 0.6988 at 35 epochs. By contrast, the two-stage adaptation reached a Macro-F1 of 0.7351 under the main adaptation protocol. Thus, longer target-only training narrowed the gap only partially. These results indicate that extended target-only training provides a limited gain on solid spectra, but it does not substitute for the gas-derived prior. The decline from 30 to 35 epochs further suggests a practical ceiling for target-only optimization on solid spectra.

Liquid spectra showed a different training-duration pattern. The target-only model increased only slightly from a Macro-F1 of 0.7198 at 15 epochs to 0.7206 at 30 epochs and 0.7290 at 35 epochs, without the decline observed for solid spectra. However, these target-only scores still remained below those obtained with gas-derived prior training. Thus, training the target-only baseline for longer did not replace the gas prior in either condensed phase, while the apparent optimization ceiling observed for solid spectra was not reproduced on liquid spectra.

We compared four routes for incorporating the gas-phase prior. The first used joint gas–target training and stopped after this stage. The second used direct gas-to-target fine-tuning, in which a model pretrained only on gas spectra was fine-tuned on the target phase. The third was the two-stage adaptation protocol defined above. The fourth was three-stage adaptation, in which a short low-rate target-phase refinement was added after the two-stage protocol.

On solid spectra, joint training and direct gas-to-target fine-tuning gave close Macro-F1 scores of 0.7212 and 0.7233. Two-stage adaptation raised the score to 0.7351. Three-stage adaptation raised it further to 0.7455. All three routes containing a target-phase fine-tuning stage outperformed joint training alone, while the two-stage and three-stage routes achieved the strongest solid-phase performance. On solid spectra, applying a target-phase fine-tuning after the joint training is therefore the more effective transfer strategy, and this additional fine-tuning is necessary to achieve the best solid performance.

On liquid spectra, the routes behaved differently. Two-stage adaptation gave a Macro-F1 of 0.7792, slightly below the 0.7817 obtained with the joint training alone. Joint training alone, direct gas-to-target fine-tuning, and three-stage adaptation reached comparable high scores of 0.7817, 0.7923, and 0.7991, with direct gas-to-target fine-tuning and three-stage adaptation nearly matched at the top. The routes could not be separated by performance on liquid spectra, which likely reflects the small size of the liquid dataset rather than a genuine equivalence of the routes. All four routes still exceeded the target-only baseline of 0.7198 by a clear margin. The gas-derived prior therefore remained beneficial on liquid spectra, while the most effective way to transfer it could not be determined from this dataset. Establishing which transfer route best suits the liquid phase remains a question for future work, one that will require a larger and more diverse liquid dataset. For this reason, the following analysis of transferability is restricted to the solid phase.

2.6. Transferability of two-stage adaptation across model families

Fig. 9 summarizes the cross-model comparison. Beyond the Dual Head architecture, we examined whether the benefit of two-stage adaptation extends to other backbones trained on the same solid-phase task. If the improvement were tied to the specific design of Dual Head, it would not survive a change of backbone. We therefore applied the identical two-stage adaptation protocol to a set of independent models. These included several capacity and structural variants within the Dual Head family, two previously published architectures for this task, IRCNN and AggMapNet, a 1D ResNet-50 backbone that we adapted to spectral input, and a self-built patch transformer. Both convolutional and attention-based architectures are included.

Fig. 9. Effect of two-stage gas-pretrained adaptation across different model families on solid-phase functional group recognition. Baseline and adapted Macro-F1 scores are shown for the Dual Head variants, published architectures, ResNet-50-1D models, and the patch transformer. Two-stage gas-pretrained adaptation improves most architectures, although the magnitude of the gain varies by model and does not remove the large performance differences among architectures.

The improvement is not specific to Dual Head. Every variant within the Dual Head family improves under adaptation, and the gain carries over to IRCNN, a convolutional model from the prior functional-group literature, as well as to the patch transformer, an attention-based architecture. That a single adaptation protocol raises performance across convolutional and attention-based backbones indicates that the gas prior supplies something the target-phase training does not, independent of how the backbone processes the spectrum.

The two ResNet backbones behave differently. Both improve under adaptation, yet their absolute Macro-F1 stays well below a usable level. We attribute this to a poor match between the residual backbone and the task rather than to any shortcoming of the adaptation, since the same protocol lifts the other backbones into a far higher range. The gas prior thus improves performance across backbones without necessarily bringing every backbone to a usable level.

AggMapNet is the single exception. It shows no gain under adaptation. Its input is not the measured spectrum but a correlation-derived 2D feature map, in which AggMap reorders the wavenumber axis by feature similarity rather than by spectral position. This departure from the measured spectral sequence may be what prevents the transfer here.

2.7. Parameter-efficient adaptation for practical use

For practical use, a condensed-phase IR recognition model should not only achieve high accuracy but also remain usable under limited computational resources.²⁹ The adapted Dual Head model used above retains detailed spectral-position information, but this also results in a large parameter count. We therefore tested whether the model could be made more parameter-efficient while still preserving the coarse spectral-location information required for functional group recognition.

To reduce the parameter count, we used a segment max-pooling strategy. The fused spectral representation was divided along the spectral axis into contiguous P segments. Within each segment, max pooling was used to retain the strongest response, and the resulting segment-level features were used for multilabel prediction. In this design, P controls the balance between compression and spectral-position resolution. Smaller P values impose stronger compression and retain only coarser location information, whereas larger P values preserve finer spectral-location structure at a higher parameter cost.

The results show a clear dependence on P (Fig. 10). When P was small, performance was substantially reduced, indicating that overly aggressive pooling removes useful positional information. As P increased, Macro-F1 improved consistently for both solid-only training and two-stage adaptation. Under solid-only training, Macro-F1 increased from 0.4968 at P = 4 to 0.6949 at P = 128. Two-stage adaptation followed the same trend, increasing from 0.6171 at P = 4 to 0.7072 at P = 32 and 0.7281 at P = 64, before remaining at a similar level at P = 128 with a Macro-F1 of 0.7248. The gas-phase models also followed this trend, rising from 0.7782 at P = 4 to 0.9053 at P = 128.

Fig. 10. Effect of segment max-pooling size P on gas-phase performance, solid-only performance, and two-stage gas-pretrained adaptation. Macro-F1 increases as P is raised from 4 to 32, after which the gains become progressively smaller and approach a plateau. Two-stage gas-pretrained adaptation consistently outperforms the corresponding solid-only baseline across all tested pooling sizes, indicating that the adaptation benefit is retained even after substantial model compression.

The parameter-performance trade-off was especially favorable at intermediate and large P values. The P = 64 model used only 2.58 million parameters, corresponding to 9.23% of the original Dual Head parameter count, yet retained 99.04% of the two-stage adaptation Macro-F1. Increasing P to 128 raised the parameter count to 4.67 million, or 16.73% of the original model, and retained 98.59% of the adapted Dual Head Macro-F1. Under solid-only training, the P = 128 model reached a Macro-F1 of 0.6949, slightly exceeding the full Dual Head solid-only baseline of 0.6902 while using only 16.73% of the parameters.

These results indicate that most of the adapted-model performance can be recovered without the full parameter count of the original Dual Head model. A very small number of spectral segments is too coarse, but once P reaches 64 or 128, the model preserves enough peak-location information to approach the full adapted model. The segment max-pooling strategy therefore provides a way to substantially reduce the parameter count of the adapted model while incurring only limited performance loss.

Condensed-phase functional-group recognition may support analytical workflows in which spectra must be interpreted rapidly or at scale. Potential applications include first-pass monitoring of functional-group changes during synthetic reactions² and batch annotation for high-throughput screening or spectral-library indexing, with related uses in quality control⁴ and materials characterization.^8,17 Such capability may be especially useful in automated or self-driving laboratories, where large numbers of spectra can be generated with limited time for manual interpretation.³⁰ Phase-aware adaptation may be particularly beneficial where labeled target-phase spectra are limited and positive examples for particular functional groups are scarce, allowing the more abundant gas-phase spectra to provide prior information beyond what can be learned reliably from the limited target-phase data alone.

2.8. Limitations and future work

This study has several limitations. The liquid-phase subset is substantially smaller than the gas- and solid-phase datasets, which limits the stability of route-level comparisons and the confidence with which phase-specific conclusions can be generalized. The condensed-phase records were contributed by many laboratories and acquired on a wide range of instrument models, but this variation was never treated as a controlled factor, so the benchmark does not support any systematic assessment of instrument-dependent behaviour. Spectra acquired in routine practice can further show baseline drift, measurement noise, varying spectral resolution, incomplete wavenumber coverage, and instrument-specific artifacts.³¹ The Min–Max normalization applied here removes a constant baseline offset by construction, but sloping or curved baselines were not corrected, and none of the remaining effects were evaluated separately. The present results therefore establish the benefit of phase-aware adaptation under the standardized preprocessing applied in this work, and they do not extend to robustness against acquisition-related distortions. In addition, the sample-preparation information recorded in the source files varies widely in terminology and level of detail, so consistent preparation categories could not be defined and each condensed-phase subset combines several distinct preparation routes. The class distribution also remains uneven, and several labels, particularly alkene and other underrepresented functional groups, continue to show limited or unstable recognition.

Future work should expand the condensed-phase benchmark with spectra collected from additional databases, instruments, and sampling conditions, with particular emphasis on liquid spectra and underrepresented labels. A subset restricted to matrix-isolation spectra would also provide a useful intermediate case between the gas phase and bulk condensed phases. Robustness should also be assessed systematically, through controlled perturbations of baseline, noise, resolution, and wavenumber coverage, and through cross-instrument evaluation. Incorporating finer-grained sample-preparation information as additional metadata or modeling variables may help account for preparation-dependent spectral variation and further improve recognition performance on liquid- and solid-phase spectra. Further progress on labels with weak or overlapping IR signatures may require complementary structural information from modalities such as NMR or mass spectrometry.^32–34 The phase-aware adaptation strategy could also be extended to broader spectral tasks, including compound classification, structural motif prediction, and multimodal spectral interpretation.

3. Conclusion

In this study, we established a phase-separated framework for multilabel functional group recognition from gas-, liquid-, and solid-phase infrared spectra. Direct transfer from gas-trained models to condensed-phase spectra was unreliable, whereas two-stage gas-pretrained adaptation substantially improved recognition on both solid and liquid spectra. The adapted model increased solid Macro-F1 from 0.6902 to 0.7351 and liquid Macro-F1 from 0.7198 to 0.7792, with the benefit extending across multiple model families. Per-class evaluation and spectral interpretation showed that the gains were chemically nonuniform, with strong recovery for nitrile but persistent difficulty for alkene when its diagnostic signatures were weakly resolved or overlapped with aromatic, saturated C–H, and fingerprint-region absorptions. Segment max pooling further reduced the model to 9.23% of the original parameter count while retaining 99.04% of its adapted Macro-F1. These results show that phase should be treated explicitly in machine-learning models for IR interpretation and that gas-phase spectra can serve as useful prior information for practical condensed-phase functional group recognition.

Author contributions

Yitian Kong conceptualized and designed the study, developed the methodology, collected and curated the data, conducted the computational experiments, validation, and formal analyses, prepared the visualizations, and wrote and revised the manuscript.

Conflicts of interest

There are no conflicts to declare.

Data availability

The infrared spectra analysed in this study are publicly available from the NIST Chemistry WebBook at https://doi.org/10.18434/T4D303. The source code and scripts used for data collection, data preprocessing, and model training are available at https://github.com/yitiankong-mk/Phase-Aware-IR. Additional data supporting the findings of this study are provided in the supplementary information (SI). Supplementary information: additional methodological details, model comparisons, statistical analyses, control experiments, and detailed results. See DOI: https://doi.org/10.1039/d6ra06513a.