Design of Supported Ionic Liquid Membranes for CO₂ Capture Using a Generative AI-Based Approach

Abstract

Growing urgency to address climate change has accelerated the development of efficient carbon capture technologies. However, traditional approaches to design materials for CO₂ capture are often hindered by time-consuming and costly experimental processes. This study investigates the application of generative AI, specifically a conditional variational autoencoder (CVAE), to accelerate the discovery and design of supported ionic liquid membranes (SILMs) for enhanced CO₂ capture. By leveraging a limited experimental data set, our CVAE model generates and predicts a large number of synthetic SILM candidates, significantly reducing the need for extensive trial-and-error experiments. The SILMs with predicted CO₂ capture capacity are then selected for synthesis and experimental evaluation. The experimental results indicate that the model demonstrates strong predictive accuracy, showing close agreement between predicted and measured values. This AI-driven approach offers a cost-effective and efficient pathway to rapidly explore vast design spaces, potentially revolutionizing the development of advanced materials for carbon capture.

1. Introduction

With energy demands at an all-time high and a heavy reliance on carbon-based fuels, concerns about the anthropogenic impact are heightened.¹ The global commitment to combating the greenhouse effect has ignited a surge in research dedicated to capturing carbon dioxide (CO₂) from the atmosphere and implementing low-carbon footprint processes—two crucial steps toward environmental protection and sustainability.² However, the development of efficient and scalable carbon capture technologies remains a significant challenge.^3,4 To capture CO₂, a wide range of materials and methods, such as cryogenic distillation, absorption, adsorption, and membrane separation, have been employed. Among these, the membrane separation is less energy-intensive. Nevertheless, the permselectivity of separation media in membrane-based processes has been a bottleneck, hindering their rapid development and large-scale deployment.^5,6

Significant research has been done on alternative approaches to overcome the limitations of membrane technology. An interesting path was the inclusion of ionic liquids (ILs) into membrane formulation, constructing supported ionic liquid membranes (SILMs). ILs are promising materials that can be tailored to be task-specific in different separation media.⁷ ILs are considered a potential alternative to traditional carbon capture systems due to their high capture capacity and selectivity,⁸ and their unique properties, such as their low vapor pressure, high thermal stability, customizable chemical structures, low flammability, and strong affinity for CO₂ absorption.⁹ ILs can be customized for specific applications by adjusting their chemistries and compositions.¹⁰⁻¹² Supporting them as in SILMs has been pursued extensively to leverage their promising properties in gas separation.^13,14 SILMs comprise a polymeric support material, typically porous, filled with an IL. The design of SILMs leverages the unique properties of ILs, such as high thermal stability and tunable physicochemical properties, important for achieving high selectivity and absorption capacities in gaseous mixtures.¹⁵ This design is beneficial for mechanical stability and allows ILs to be effectively utilized in demanding industrial environments. The performance of SILMs in CO₂ capture is influenced by the IL selection, the choice of the support material, and membrane morphology. These factors are crucial as they regulate the interaction strength and selectivity toward CO₂, optimizing the membrane for specific separation targets and operational conditions. Since, the choice of host polymer and IL greatly influences the gas separation performance, offering a path to improve gas transport properties and the long-term stability of the membrane.¹⁵ Along these lines, Friess et al. reported that semicrystalline fluoropolymers outperform other polymer supports in separating CO₂/N₂ mixtures.¹³ This superior gas separation performance in fluoropolymers was attributed to the unique compositional blend between the crystalline and amorphous phases of the polymer, which offers improved permeability and adequate selectivity.

While SILMs have shown promising results, particularly for capturing CO₂, there remains a significant challenge in finding the optimal SILM composition, fine-tuned for a specific feed composition.^16,17 The expensive and labor-intensive process of searching for novel materials is still a drawback, slowing down research in SILMs. Given the rapid advancements of artificial intelligence (AI), it is now feasible to accelerate the wet lab experiments and arrive at optimum SILM compositions for capturing CO₂ from different streams. While AI frameworks have been successfully employed to empower material selection tailored for other purposes, e.g., batteries and gas sorption,^18,19 the main barrier of data limitation still exists for material synthesis.²⁰

Recently, generative AI has emerged as a promising strategy for material design. This algorithmic strategy can effectively expand limited experimental data sets, reducing the time and resources required for costly laboratory work. Moreover, generative AI enables the exploration of hypothetical conditions that would be challenging to achieve experimentally, such as extreme temperatures or pressures. The recent surge in utilization of generative AI largely stems from the advent of large language models, which have catalyzed new applications across various fields, including chemistry. Figure 1 presents an overview of the AI paradigms and their roles in chemical knowledge. While generative AI applications range from drug discovery^21,22 to process design²³ and multifunctional material synthesis,²⁴ its use for SILMs has not been explored.

Figure 1.

Schematic representation of AI paradigms in chemistry. Discriminative AI (top) performs classification and regression tasks for molecular property prediction, while generative AI (bottom) creates new molecules, synthesis routes, and materials by learning from existing chemical data distributions.

In this study, we adopted a machine learning (ML) strategy within a generative AI framework to simulate the performance of SILMs in CO₂ capture. Using this surrogate model, we generated artificial SILMs based on a limited set of synthesized ones, which aids in the design of new SILMs for CO₂ capture. We present a systematic approach combining experimental synthesis and AI-guided design of SILMs. First, we synthesized SILMs and evaluated their CO₂ sorption performance under various operating conditions. This experimental data set was used to train a conditional variational autoencoder (CVAE) as our generative AI framework. The trained model generated new SILM candidates, from which we selected the most promising SILMs based on predicted sorption capacity. Finally, we synthesized these AI-proposed SILMs and experimentally validated their CO₂ absorption performance against model predictions to assess the effectiveness of our AI-guided design approach. Figure 2 provides a schematic overview of our methodology, representing the use of generative-AI to guide the design of new SILMs. The organization of the rest of this paper is as follows: Section 2 details the experimental and computational methodologies, including the synthesis of SILMs, the CO₂ sorption testing methods, and the development of the CVAE ML model. In Section 3, we assess the model’s performance in generating new SILMs and evaluate its effectiveness in the inverse design of SILMs. Finally, Section 4 presents some concluding remarks.

Figure 2.

Process of using generative AI to design optimal SILMs for CO₂ absorption. The workflow begins with fabricating a limited number of SILMs and conducting CO₂ sorption tests, followed by data preprocessing. A conditional variational autoencoder is then used to generate new SILM candidates, which are evaluated and experimentally validated for their CO₂ absorption performance to determine the effectiveness of the generative AI approach.

2. Methodology

2.1. Experimental Materials

The experimental work involved obtaining poly(vinylidene fluoride) [PVDF] of different molecular weights (M_W): approximately 180, 352, 275, 534, and 1000 kg/mol, as well as a random block copolymer, Poly(vinylidene fluoride-co-hexafluoropropylene) [PVDF-HFP], with molecular weights around 400, and 455 kg/mol from Sigma-Aldrich (St. Louis, MO). PVDF (Kynar) and PVDF-HFP (255 kg/mol) were obtained from PolyK (PA). PEBAX block copolymers with different grades, MH-1657, 2533-SA-01-MED, and RNEW (30R51-SA-01), were procured from ARKEMA (King of Prussia, PA). Imidazolium-based IL, 1-ethyl-3-methylimidazolium bis(trifluoromethyl sulfonyl)imide (EMIM[TF₂N], Purity ≥98%), Triethyl phosphate (TEP, ≥99%), and Dimethylacetamide (DMAc, >99%) were sourced from Sigma-Aldrich (St. Louis, MO). Instrument-grade carbon dioxide gas (CO₂) exceeding 99.99% purity was obtained from Matheson Trigas, Inc. (Lincoln, NE). To estimate CO₂ sorption, we used our developed method detailed in our previous work.²⁵

2.2. SILM Synthesis

To investigate CO₂ absorption in SILMs, first, a series of polymeric mixtures, dope solutions consisting of polymer, IL, and solvent, were prepared. To prepare PVDF-based SILMs, polymer dope solutions were prepared by dissolving the polymer in DMAc through continuous mixing of the solution at 95 °C for 24 h. For PVDF-HFP based SILMs, the polymer dope solution was prepared in TEP using the same condition noted above. Similarly, for PEBAX-based SILMs, the dope solution for grade (1657) was prepared in a mixture of 30:70 v/v % water in ethanol. For grade (2533), ethanol was used as the solvent, and for the grade (RNEW), we used a mixture of 3:1 v/v of 1-Butanol in 1-Propanol. All mixtures were prepared by dissolving the polymer in a solvent and continuously stirring the mixture at 95 °C for 24 h. Then, the mixtures were allowed to cool to ambient temperature. SILM mixtures of equal wt % (50:50) of host polymer and ionic liquid EMIM [TF₂N] were then prepared. This was achieved by adding a stock solution of the polymeric mixture to 6-dram capped glass vials (VWR, Radnor, PA, Borosilicate Glass Vials) and introducing a predetermined quantity of ionic liquid to each solution to achieve the intended composition. To achieve a homogeneous solution, the mixtures were further stirred and heated using a hot plate manufactured by Heidolph Inc. (Wood Dale, IL). This process was carried out for 24 h before the solution was used for spin coating. For CO₂ absorption analysis, thin films of SILMs were directly deposited on quartz crystal (QC) substrates via spin-coating. The specific spin-coating conditions are detailed in the Supporting Information (SI), section 1.4.

2.3. CO₂ Sorption Measurement

The sorption isotherms for SILMs were measured using a quartz crystal microbalance (QCM). The linear correlation between the frequency shift and the mass loading, described by the Sauerbrey equation, was used to measure the mass of CO₂ absorbed within the polymeric film. Before coating the QCM substrates, they were cleaned using multistep solvent cleaning as described elsewhere.¹⁷ To coat a thin film of SILM on the QC, 100 μL of the dope solution, 50:50 wt % polymer to IL, was dispensed onto QC substrates, followed by spin coating. The mass of the resulting film was measured using an analytical balance (Sartorius, Bohemia, NY, MSA225P100DI Cubis Analytical Balance, which has a precision of ±0.01 mg) before and after the coating process. The prepared samples were then placed into the QCM holder for gas sorption measurement. The QCM holder was inserted into a high-pressure sealed chamber. Briefly, the system was evacuated using a rotary vane vacuum pump, and the temperature of the chiller was set. Once the coated substrate had reached a stable frequency, the Eon-LT software was used to record the frequency and temperature of the QCM module. The pressure inside the chamber was measured using a pressure sensor (Honeywell, PX2AN1XX200PSAAX) and transmitted to LabVIEW via microprocessors. Subsequently, the pressure was regulated using a pressure regulator, and CO₂ was introduced into the system. The CO₂ sorption measurements were conducted at six different pressures of 50, 100, 125, 150, 175, and 200 psi and four different temperatures of 10, 20, 30, and 40 °C.

2.4. AI Generation of SILMs Using CVAE

After completing the experimental work, we evaluated and preprocessed our data set to ensure its reliability and accuracy as input for our ML model. The preprocessing involved two key steps: eliminating outliers and averaging repeated measurements. It is generally observed that extracting patterns from a data set with a distribution close to normal is more effective for ML model development.²⁶ Therefore, our preprocessing aimed to bring the data closer to a normal distribution, as illustrated in Figure 3. As shown, the quantile–quantile (Q–Q) plot demonstrates the improved alignment of our data with a normal distribution, while the box plot shows a notable reduction in outliers. For each unique combination of host polymers, molecular weight, temperature, and pressure, we averaged the CO₂ absorption capacity across four separate measurements to mitigate experimental variations. These steps resulted in a refined data set of 216 unique data points, encompassing various host polymers, molecular weights, temperatures, and pressures. While this data set provides valuable insights into CO₂ absorption behavior in SILMs, its relatively small size—given the extensive range of possible SILM compositions and operating conditions—limits its effectiveness for the design of SILMs for CO₂ capture across diverse scenarios. This limitation motivated the development of our generative AI framework for creating novel SILMs, enabling us to explore a broader range of potential SILM compositions and predict their CO₂ capture performance, thereby overcoming the constraints of our experimental data set.

Figure 3.

Statistical analysis of CO₂ absorption data in SILMs before and after preprocessing: Distribution (A, D), Q–Q plots (B, E), and box plots (C, F).

To develop a predictive model, a CVAE was selected to guide the experimental work in the synthesis of SILM due to its ease of training, stability, and the advantage of conditional generation. Conditional generation enables the model to create synthetic data while maintaining control over specific input parameters (in this case, temperature, pressure, host polymer type, and molecular weight), ensuring the generated samples align with desired experimental conditions. The CVAE’s ability to generate samples conditioned on specific attributes makes it particularly suitable for exploring the vast space of potential SILM compositions while targeting desired properties. As illustrated in Figure 4, CVAE framework implementation involves three stages to effectively generate synthetic data that mirrors the statistical properties of the original data set. A key stage is the encoder, a neural network tasked with converting the input data and its associated conditional information into a latent space (z). In this process, the input data of SILM, representing the CO₂ sorption capacity of SILMs, is concatenated with conditions including molecular weight (M_W), host polymer, pressure, and temperature, and then fed into a fully connected neural network. The host polymer is converted into a machine-readable format using one-hot encoding, allowing us to analyze its effect on CO₂ absorption. The encoder network transforms the data into parameters of a latent space distribution, outputting a mean vector (μ) and a log-variance vector (log σ²).

Figure 4.

Representation of the architecture of the CVAE for SILM generation. The encoder compresses experimental SILMs data into a latent space; after sampling from the latent space, the decoder reconstructs new SILM designs from the latent space. This structure enables the model to learn and reproduce complex SILM properties for optimized CO₂ absorption.

1

2

Equation 1 defines the posterior distribution of the latent variables (z) given the input SILM data (x) and conditions (c). This distribution is assumed to be Gaussian with mean μ and variance σ². Equation 2 shows how the encoder function (f_encoder) processes the input SILM data and conditions to produce the mean and log-variance of the latent space distribution. Together, these equations enable the model to compress the complex SILM properties into a lower-dimensional latent space.

In the second stage, we employ the reparameterization approach to sample from the latent space. This allows for gradient backpropagation through the sampling process, which is essential for training the model. The sampling is performed as:

3

Equation 3 demonstrates how we generate a sample z from the latent space. Here, ε is a random noise vector drawn from a standard normal distribution, and ⊙ represents element-wise multiplication. This reparameterization allows the model to explore the latent space of SILM properties while maintaining differentiability, which is crucial for training the model through backpropagation.

The third stage involves the decoder, which reconstructs the input data from the latent variables and the conditional information. The decoder takes the latent variables z and the conditions c, concatenates them, and processes them through dense layers to generate the output data x̂. The decoding process is represented as:

4

5

Equation 4 represents the probability distribution of the reconstructed SILM data (x̂) given the latent variables (z) and conditions (c). Equation 5 shows how the decoder function (f_decoder) generates the output SILM data x̂ from the latent variables z and the conditions c. Using this approach, SILM properties are reconstructed from the compressed latent representation, allowing the model to generate new SILM designs with desired CO₂ absorption characteristics.

After constructing the model, the CVAE was optimized to minimize a loss function consisting of two components. The first component is the reconstruction loss, implemented as Mean Squared Error (MSE), which measures how well the model can reconstruct the input data. The second component is the KL (Kullback–Leibler) divergence, which measures how much the learned latent variable distribution deviates from the prior distribution. Since this specific combination of loss components is not directly available in TensorFlow, we implemented the custom loss function as follow:

6

The MSE (x, x̂) term represents the reconstruction loss between the original input x and the reconstructed output x̂. This term quantifies how well the model reconstructs the input data from the latent variables and conditions. Similarly, the term KL(q(z|x,c)∥p(z|c)) measures how much the learned latent variable distribution deviates from the prior distribution. This term is minimized to ensure the latent space is smooth and continuous, aligning the learned distribution with the prior distribution. A detailed explanation of the CVAE optimization can be found elsewhere.²⁷

3. Results and Discussion

3.1. CVAE Training and SILMs Generation

The CVAE model is trained using the Adam optimizer, and the training process is set to run for a maximum of 1500 epochs, with a batch size of 32. A 20% portion of the data is reserved for validation. The model aims to minimize a custom loss function that combines the reconstruction error and the Kullback–Leibler divergence.²⁸ Early stopping is implemented with a patience of 150 epochs, meaning the training will stop if no improvement in validation loss is observed for 150 consecutive epochs. This helps prevent overfitting and unnecessary computation. The best weights are restored when early stopping is triggered. The model architecture includes two dense layers in the encoder (64 units in the first hidden layer, 32 units in the second hidden layer, followed by output layers for mean and log variance) and two dense layers in the decoder (32 units in the first hidden layer, 64 units in the second hidden layer, followed by the output layer), with ReLU activation functions. Dropout layers with a rate of 0.2 are employed after each dense layer except the output layers. L₁ and L₂ regularization (both with a rate of 0.001) are applied to the dense layers to further prevent overfitting. The latent space dimension is set to 2, allowing for a balance between data compression and reconstruction quality.

Figure 5 illustrates the training process and performance of our CVAE for reconstructing of CO₂ sorption distribution in SILMs at different epochs. The error loss during the epochs in the training of the CVAE framework for generating the SILMs data is shown in panel A. The smooth curves of the training loss and validation loss suggest that the ML model has successfully learned the underlying patterns defining SILM's properties. The consistent decrease in both training loss and validation loss indicates good generalization and a well-regulated training process. The close proximity of the two curves throughout the training suggests that the model is not overfitting, maintaining its ability for generalization. The gradual flattening of both curves toward the later epochs indicates that the model is approaching convergence, having captured the essential features of the SILM data distribution.

Figure 5.

CVAE training for CO₂ absorption in SILMs. (A) Loss curves over 800 epochs. (B–G) Evolution of AI-generated CO₂ absorption distribution (orange) compared to actual SILM data (blue) at epochs 1, 5, 10, 50, 100, and 200.

The loss curve demonstrates the model’s learning progression over the first 800 epochs, with training and validation losses converging, further confirming good generalization. We highlight key epochs (1, 5, 10, 50, 100, 200) to show the rapid initial improvement and subsequent refinement of the model. Panels B–G showcase the evolution of the AI-generated CO₂ sorption distribution compared to actual SILM data across these selected epochs. At epoch 1 (panel B), the generated distribution is narrow and misaligned with the actual data. By epoch 5 (panel C), the distribution begins to spread, and epoch 10 (panel D) shows further refinement in shape. Significant improvement in matching the actual distribution is evident by epoch 50 (panel E). Epoch 100 (panel F) demonstrates a close alignment with the experimental data, and finally, at epoch 200 (panel G), the AI-generated distribution closely mirrors the experimental SILM data. This progression indicates successful learning and accurate representation of CO₂ absorption characteristics in SILMs by our CVAE model.

After training the CVAE framework, the model generated 1000 artificial SILMs and their probable CO₂ capture capacities under various thermodynamic conditions. Figure 6 demonstrates the CVAE model’s ability to capture key distributions of SILM properties with varying degrees of alignment across different parameters. Each figure shows the statistical distribution of each input parameter, comparing the original data with the synthesized data generated by the ML model. The temperature (A) and pressure (B) distributions show strong agreement between actual and AI-generated SILMs, indicating the model’s proficiency in replicating thermodynamic conditions. The CO₂ absorption distribution (D) also shows good alignment, particularly in the peak and overall shape, suggesting the model has effectively learned the absorption behavior. Notably, there is a discrepancy in the molecular weight distribution (C). While the AI-generated distribution captures the general trend, it shows a broader and flatter distribution, particularly at higher molecular weights. This difference can be attributed to two factors: first, the limited number of discrete molecular weights in our experimental training data set, and second, the model’s exploration of a wider molecular weight range that could represent new opportunities for SILM design. The fact that the model maintains correlation patterns with CO₂ absorption (−0.22 in correlation heatmaps) across this expanded range suggests these unexplored molecular weight regions merit further experimental investigation. The host polymer distribution (E) demonstrates the model’s ability to generate a balanced representation across polymer types. Further statistical comparison between actual and AI-generated SILMs is provided in SI, section 4.

Figure 6.

(A) Comparison of temperature distribution between actual and AI-generated SILMs, showing close alignment in the range and frequency of operating temperatures. (B) Pressure distribution comparison illustrates similar pressure condition patterns used for both real and AI-generated SILMs. (C) Molecular weight distribution overlay, demonstrating the AI model’s ability to generate SILMs with molecular weights consistent with actual data. (D) CO₂ absorption capacity distribution, highlighting the similarity in absorption properties between real and AI-generated SILMs. (E) Host polymer distribution bar chart, comparing the frequency of different polymers used in actual SILMs versus those generated by the AI model.

The correlation heatmaps in Figure 7 provide further insight into the CVAE model’s performance in capturing relationships between variables. The model successfully replicates the strong positive correlation between pressure and CO₂ absorption (0.83 actual, 0.87 AI-generated) and the moderate negative correlation between temperature and absorption (−0.41 actual, −0.39 AI-generated). These similarities demonstrate the model’s grasp of fundamental thermodynamic principles governing CO₂ absorption. The correlation between molecular weight and CO₂ absorption shows a similar pattern (−0.22 actual, −0.22 AI-generated). This agreement suggests that the model accurately captures this relationship within the explored parameter space. Despite minor differences in correlations, the overall similarity in correlation patterns indicates that the AI-generated SILMs largely mimic the complex interplay of variables observed in actual SILMs, while also suggesting new avenues for experimental investigation in SILM design and optimization. A SHAP analysis to investigate the most influential parameters affecting CO₂ absorption in AI-generated SILMs is included in the SI, section 2. This analysis reveals the relative importance and impact range of various parameters on CO₂ absorption predictions.

Figure 7.

Correlation heatmaps for actual (A) and AI-generated (B) SILMs. The heatmaps illustrate the relationships between temperature, pressure, molecular weight, and CO₂ absorption capacity.

To examine our CVAE model’s predictive capabilities, we compared the experimental and AI-generated parameter ranges in Table 1. Table 1 summarizes the operating ranges for temperature (10–40 °C vs 2.56–47.5 °C), pressure (50–200 psi vs 12.76–237.47 psi), and molecular weights for different polymer types, comparing experimental data with CVAE predictions. The comparison demonstrates that while the model encompasses the experimental ranges, it also reasonably explores beyond these boundaries while maintaining physically meaningful predictions.

Table 1. Comparative Analysis of Experimental and CVAE Ranges for Temperature, Pressure, Molecular Weight, and Host Polymer Types.

host polymer	molecular weights [exp.] in kg/mol	molecular weights [CVAE] in kg/mol	temp. range (°C) [exp.]	temp. range (°C) [CVAE]	pressure range (psi) [exp.]	pressure range (psi) [CVAE]
PVDF	180–2000	10–2500	10–40	2.5–47.5	50–200	12–237
PVDF-HFP	244–455		10–40		50–200
PEBAX	100–240		10–40		50–200

After expanding the available sample of SILMs using the CVAE model trained on a limited number of actual SILMs, we selected the best candidates based on their predicted CO₂ absorption performance. Table 2 presents a comparison of the top 3 SILMs, including both experimentally synthesized samples and those generated by the model. This comparison allows us to evaluate the potential of AI-generated SILMs against known, high-performing actual SILMs. Notably, the CVAE model demonstrated a remarkable ability to capture and leverage the underlying thermodynamic principles governing CO₂ sorption. As shown in Table 2, the model consistently proposed SILMs with lower operating temperatures (5.7 to 8.3 °C) and maintained high pressures (229 to 235 psi), aligning with established thermodynamic principles favoring gas sorption. Although our original data did not include the specific pressure and temperature ranges proposed by the CVAE model, the model was able to distinguish the pattern of favorable pressure and temperature in CO₂ sorption. The model’s ability to understand the thermodynamic conditions under which SILMs exhibit high CO₂ sorption highlights the value of using an AI framework. Properties like host polymer and molecular weight are important, as their effects on CO₂ sorption are not thoroughly examined and can be better deciphered using data-driven methods.

Table 2. Comparison of CO₂ Sorption Data: Actual Synthesized SILMs vs AI-Generated SILMs.

	thermodynamic properties		material properties				thermodynamic properties		material properties
	temperature (°C)	pressure (psi)	MW (g/mol)	host polymer	moles of CO2/kg sorbent		temperature (°C)	pressure (psi)	MW (g/mol)	host polymer	predicted moles of CO2/kg sorbent	actual moles of CO2/kg sorbent	relative error (%)
actual synthesized SILMs	10	200	534,000	PVDF	1.228 ± 0.200	AI-generated SILMs	8.3	229	454,544	PVDF-HFP	1.35	1.547 ± 0.059	12.74
	10	200	400,000	PVDF-HFP	1.27 ± 0.101		7.3	235	390,130	PVDF	1.32	1.298 ± 0.475	1.69
	10	175	534,000	PVDF	1.066 ± 0.370		5.7	229	251,701	PVDF-HFP	1.31	1.302 ± 0.109	0.15

To validate our AI-driven approach, we synthesized the top-performing AI-generated SILM candidates in the laboratory following the same protocols used for our original experimental samples. Subsequent CO₂ absorption measurements on these AI-guided SILMs revealed promising results, with several candidates demonstrating absorption capacities comparable to or exceeding those of the best-performing actual SILMs. This outcome not only validates the predictive power of our CVAE model but also highlights its potential as a powerful tool for materials discovery. The success of these AI-guided syntheses underscores the practical utility of generative AI framework in the quest for optimal materials. By leveraging a limited number of initial experiments to generate and identify promising SILM candidates, this approach offers a pragmatic solution to the challenge of efficient material discovery.

Table 2 also compares the CO₂ sorption data of actual synthesized SILMs with AI-generated SILMs. To ensure the model’s performance, the predicted CO₂ absorption by the AI-generated model was cross-validated with experimental methods. This involved examining SILMs synthesized with PVDF and PVDF-HFP grades with the nearest molecular weights available commercially compared to AI-generated host polymers under their respective thermodynamic conditions. For instance, SILM synthesized with PVDF-HFP (455 kg/mol), the predicted absorption capacity of 1.35 mol CO₂/kg sorbent at 8.3 °C and 229 psi closely matched the actual measured value of 1.547 ± 0.059 mol CO₂/kg sorbent at given thermodynamic condition. Similarly, for PVDF (352 kg/mol), the predicted sorption rate of 1.32 CO₂/kg sorbent at 7.3 °C and 235 psi was in line with the tested measurement of 1.298 ± 0.475 mol CO₂/kg sorbent. For PVDF-HFP (244 kg/mol), the predicted value of 1.31 CO₂/kg sorbent at 5.7 °C and 229 psi was consistent with the actual measurement of 1.302 ± 0.109 mol CO₂/kg sorbent. These results highlight the robustness of the CVAE model in forecasting SILM performance with a high degree of accuracy. The close agreement between predicted and actual values underscores the model’s reliability and its potential to accelerate the discovery and optimization of high-performing CO₂ separation materials. A detailed analysis on the role of host polymer on CO₂ sorption capacity in SILMs is discussed in SI, section 3.

4. Conclusions

This study demonstrates the significant potential of generative AI (specially autoencoders) in guiding the design of SILMs for CO₂ capture. With limited experimental data, our CVAE model successfully generated novel SILM candidates with enhanced CO₂ sorption capacities. The model’s ability to identify the suitable operating conditions and explore a wider range of material properties underscores its effectiveness in capturing complex relationships between SILM characteristics and CO₂ sorption performance. Our CVAE model successfully predicted high-performing SILM candidates, with predictions closely aligning with experimental results. For instance, the predicted CO₂ absorption capacities for SILMs synthesized with PVDF-HFP (455 kg/mol), PVDF (352 kg/mol), and PVDF-HFP (244 kg/mol) were 1.35, 1.32, and 1.31 mol CO₂/kg sorbent, respectively, compared to experimental values of 1.547 ± 0.059, 1.298 ± 0.475, and 1.302 ± 0.109 mol CO₂/kg sorbent. This close agreement demonstrates the model’s robustness and accuracy, highlighting its potential to accelerate material discovery and design. Our work establishes a new paradigm for integrating machine learning with materials science, offering a more efficient and targeted approach for addressing critical environmental challenges like carbon capture.

Data Availability Statement

The data supporting this study’s findings are available from the corresponding author upon reasonable request.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.iecr.4c03280.

• Section 1 includes general information regarding instrumentation used during experimental and characterization procedures, including Fourier-Transform Infrared Spectroscopy (FT-IR), Polarized Light Microscopy, X-ray Diffraction (XRD), and Spin Coater. Section 2 includes SHAP analysis for evaluating the impact of key parameters on CO₂ absorption predictions in SILMs, highlighting the relative importance of temperature, pressure, molecular weight, and host polymer type. Section 3 includes the influence of role of host polymer on the CO₂ capturing capacity of SILMs. Section 4 includes statistical analysis of parameters (PDF)

Author Contributions

^† S.I. and H.S. contributed equally to this work. Conceptualization: M.B. Methodology: S.I., H.S., and M.B. H.S. developed the CVAE ML model. Investigation: S.I., H.S., and M.B. Funding acquisition: M.B. Project administration: M.B. Supervision: M.B. Writing: S.I. and H.S. Writing-reviewing and editing; M.B. All authors have read and approved the final manuscript.

The authors declare no competing financial interest.

Special Issue

Published as part of Industrial & Engineering Chemistry Researchspecial issue “Inverse Design of Materials and Processes for Separations”.