Conductive aerogels, ultralight materials with adjustable mechanical properties and excellent electrical performance, hold significant promise for diverse applications, including pressure sensors, electromagnetic interference shielding, thermal insulation, and wearable heaters. Traditional fabrication methods involve iterative, time-consuming experimentation across a vast parameter space to optimize the ratios of conductive nanomaterials (e.g., carbon nanotubes, graphene, Ti3C2Tx MXene nanosheets), functional fillers (e.g., cellulose nanofibers, silk nanofibrils, chitosan), polymeric binders (e.g., gelatin), and crosslinking agents (e.g., glutaraldehyde). The complex interplay between composition, structure, and properties in these aerogels makes achieving user-specified mechanical and electrical properties a challenging and labor-intensive process. This necessitates a significant number of iterative experiments to identify the optimal fabrication parameters, which is both time-consuming and resource-intensive. Machine learning (ML), a branch of artificial intelligence (AI), offers a powerful approach to overcome this challenge. AI/ML techniques are well-suited for unraveling intricate correlations within complex parameter spaces. Their application in materials science has grown rapidly, particularly in areas with readily available simulation programs and high-throughput analytical tools that generate extensive datasets. However, applying AI/ML to conductive aerogel design faces significant hurdles, mainly due to the scarcity of high-quality data points. This scarcity stems from the lack of standardized fabrication protocols, variations in the building blocks employed across different research labs, and the time-consuming nature of aerogel fabrication. Most prior work focuses on optimizing a single property, neglecting the complex interactions between different properties. The development of a predictive model capable of efficiently suggesting the ideal parameter set for a conductive aerogel with multiple, programmable properties would drastically accelerate the design process.

The literature extensively covers the fabrication and properties of conductive aerogels based on various materials and fabrication techniques. Numerous studies have explored the use of different nanomaterials like carbon nanotubes, graphene, and MXene to enhance electrical conductivity. The incorporation of functional fillers such as cellulose nanofibers, silk nanofibrils, and chitosan have been shown to improve mechanical properties, particularly flexibility and strength. The use of polymeric binders and crosslinking agents play a crucial role in controlling the aerogel's structure and overall stability. However, a comprehensive understanding of the complex interactions between these components and their influence on the overall properties of the aerogel remains largely unexplored. Most studies focus on optimizing individual properties, often neglecting the intricate relationships between mechanical and electrical properties. This research gap highlights the need for a more holistic and efficient approach to conductive aerogel design.

This study presents an integrated platform combining collaborative robotics and AI/ML to accelerate the design of conductive aerogels. The methodology consists of three primary phases: (1) establishing a feasible parameter space; (2) implementing active learning loops; and (3) synthesizing virtual data points. Phase 1 utilizes an automated pipetting robot (OT-2) to prepare 264 mixtures of Ti3C2Tx MXene, cellulose nanofibers (CNFs), gelatin, and glutaraldehyde (GA) with varying ratios and mixture loadings. These mixtures are freeze-dried to produce aerogels, which are then categorized based on their structural integrity (A-grade, B-grade, C-grade). A support vector machine (SVM) classifier is trained to predict the likelihood of obtaining A-grade aerogels based on the fabrication parameters. This classifier helps define a feasible parameter space for further investigation. Phase 2 involves 8 active learning loops, where the AI/ML model guides the experimental design. In each loop, the OT-2 robot prepares mixtures based on parameters identified as least familiar to the current model. A UR5e robotic arm integrated with an Instron compression tester automates aerogel characterization, significantly increasing data acquisition rates. The data from the compression tests (compressive stress at 30% strain, σ30), electrical resistance measurements (initial electrical resistance, R0) are combined with the fabrication parameters to train an artificial neural network (ANN) model. Phase 3 employs data augmentation to improve the ANN model's accuracy. The User Input Principle (UIP) method synthesizes virtual data points based on observed experimental variations. The combination of real and virtual data points enhances the model's training, mitigating overfitting. The ANN model's performance is evaluated using mean absolute error (MAE) for compressive stress and mean relative error (MRE) for electrical resistance. The model's interpretability is enhanced through SHapley Additive exPlanations (SHAP) analysis. Finite Element (FE) simulations are conducted to validate the correlations between mixture loading, density and compressive strength. The final stage involves using the trained ANN model to perform two-way design tasks. First, it predicts the mechanical and electrical properties of aerogels based on specified fabrication parameters. Second, it inversely designs aerogels to meet specific property requirements (e.g., high strength, low resistance). This automated design process eliminates the need for extensive iterative experimentation.

The integrated robotics/ML platform successfully generated a high-accuracy prediction model for conductive aerogel properties. The SVM classifier effectively defined a feasible parameter space, filtering out combinations leading to fragile aerogels. The active learning approach, combined with data augmentation, significantly increased the data acquisition rate and improved the ANN model's predictive accuracy. The MAE for compressive strength decreased from 2.5 to 1.5 kPa and the MRE for electrical resistance decreased from 49% to 18.4% throughout the 8 active learning loops. The ANN model outperformed other machine learning models (linear regression, decision tree, gradient boosted decision tree, random forest). SHAP analysis revealed that mixture loading (and therefore density) had the most significant impact on compressive strength, while MXene loading most strongly influenced electrical resistance. FE simulations validated the experimentally observed relationship between density and compressive strength. The model successfully predicted and inversely designed aerogels to meet specific property targets. Finally, the research identified a strain-insensitive conductive aerogel (MXene/CNF/gelatin/GA ratio of 78/13/9/-, mixture loading of 7.5 mg mL-1) exhibiting high conductivity (R0 < 20 Ω), suitable compressive strength (σ30 = 4.0 kPa), and ultralow pressure sensitivity (0.02 kPa-1), making it highly suitable for wearable thermal management applications. This aerogel showed stable Joule heating performance under repetitive compression cycles, reaching temperatures up to 70 °C at 2.0 V within 300 s.

This study's findings significantly advance the design and development of conductive aerogels. The integrated robotics/ML approach offers a substantial improvement over traditional methods, reducing experimental time and resources while improving the accuracy of property prediction and enabling the inverse design of aerogels with custom-tailored properties. The high accuracy of the prediction model across the entire parameter space allows for efficient exploration of the design space and rapid response to changing design requirements. The integration of SHAP analysis and FE simulations provides a deeper understanding of the structure-property relationships in these complex materials, moving beyond a "black box" model to provide data-driven design principles. The successful demonstration of a strain-insensitive conductive aerogel for wearable thermal management showcases the practical applications of this approach. The methodology developed is not limited to conductive aerogels and could be adapted to other materials science applications where high-throughput experimentation and accurate property prediction are crucial.

This research presents a novel integrated workflow combining automated robotic experiments, AI/ML prediction modeling, and FE simulations for the design of conductive aerogels with programmable properties. The resulting high-accuracy prediction model enables both forward and inverse design tasks, significantly accelerating the development process. The identification of a strain-insensitive conductive aerogel suitable for wearable thermal management highlights the practical impact of this approach. Future research could explore expanding the range of materials and properties included in the model, investigating more complex aerogel architectures, and further enhancing model interpretability.

While the study demonstrates significant advancements, some limitations exist. The current model focuses on a specific set of materials and fabrication parameters. Extending the model to include a wider range of materials and processing techniques would enhance its generality. The accuracy of the model is dependent on the quality of the experimental data. Careful attention to experimental design and data quality control is crucial. The FE simulations simplified the aerogel microstructure, potentially neglecting certain aspects of the real material's behavior. Further refinement of the FE models could improve the accuracy of the simulations.