Vitrimers represent an emerging class of sustainable polymers with self-healing capabilities enabled by dynamic covalent adaptive networks. However, their limited molecular diversity constrains their property space and potential applications. Recent developments in machine learning (ML) techniques accelerate polymer design by predicting properties and virtually screening candidates, yet the scarcity of available experimental vitrimer data poses challenges in training ML models. To address this, we leverage molecular dynamics (MD) data generated by our previous work to train and benchmark seven ML models covering six feature representations for glass transition temperature (Tg) prediction. By averaging predicted Tg from different models, the model ensemble approach outperforms individual models, allowing for accurate and efficient property prediction on unlabeled datasets. Two novel vitrimers are identified and synthesized, exhibiting experimentally validated higher Tg than existing bifunctional transesterification vitrimers, along with demonstrated healability. This work explores the possibility of using MD data to train ML models in the absence of sufficient experimental data, enabling the discovery of novel, synthesizable polymer chemistries with a wide range of desirable properties. The integrated MD–ML approach offers polymer chemists an efficient tool for accurate property prediction and designing polymers tailored to diverse applications.

Spinodal metamaterials, with architectures inspired by natural phase-separation processes, have presented a significant alternative to periodic and symmetric morphologies when designing mechanical metamaterials with extreme performance. While their elastic mechanical properties have been systematically determined, their large-deformation, nonlinear responses have been challenging to predict and design, in part due to limited data sets and the need for complex nonlinear simulations. This work presents a novel physics-enhanced machine learning (ML) and optimization framework tailored to address the challenges of designing intricate spinodal metamaterials with customized mechanical properties in large-deformation scenarios where computational modeling is restrictive and experimental data is sparse. By utilizing large-deformation experimental data directly, this approach facilitates the inverse design of spinodal structures with precise finite-strain mechanical responses. The framework sheds light on instability-induced pattern formation in spinodal metamaterials—observed experimentally and in selected nonlinear simulations—leveraging physics-based inductive biases in the form of nonconvex energetic potentials. Altogether, this combined ML, experimental, and computational effort provides a route for efficient and accurate design of complex spinodal metamaterials for large-deformation scenarios where energy absorption and prediction of nonlinear failure mechanisms is essential.

The thermal conductivity of covalent organic frameworks (COFs), an emerging class of nanoporous polymeric materials, is crucial for many applications, yet the link between their structure and thermal properties remains poorly understood. Analysis of a dataset containing over 2400 COFs reveals that conventional features such as density, pore size, void fraction, and surface area do not reliably predict thermal conductivity. To address this, an attention-based machine learning model was trained, accurately predicting thermal conductivities even for structures outside the training set. The attention mechanism was then utilized to investigate the model's success. The analysis identified dangling molecular branches as a key predictor of thermal conductivity, leading us to define the dangling mass ratio (DMR), a descriptor that quantifies the fraction of atomic mass in dangling branches relative to the total COF mass. Feature importance assessments on regression models confirm the significance of DMR in predicting thermal conductivity. These findings indicate that COFs with dangling functional groups exhibit lower thermal transfer capabilities. Molecular dynamics simulations support this observation, revealing significant mismatches in the vibrational density of states due to the presence of dangling branches.

Deep learning has revolutionized scientific and engineering applications by enabling fast, data-driven predictions and optimizations. In materials science, however, its impact is limited by complex structure–property relationships, sparse high-quality data, and the need to respect fundamental physical laws. Overcoming these challenges calls for machine learning frameworks that learn effectively from limited data while producing results that are physically meaningful, interpretable, and actionable.
This thesis develops physics-guided machine learning approaches to accelerate the design, modeling, and interpretation of materials across scales and material classes. It introduces neural network architectures for metamaterial design that learn directly from data while remaining physically consistent. These models retrieve structural designs that achieve specified target properties instantly, enabling rapid exploration of the structure–property landscape and supporting scenarios with multiple design goals more efficiently than traditional optimization methods. For material modeling, this work addresses the limitation that stress fields are not directly accessible in experiments. The proposed frameworks instead learn constitutive laws from measurable quantities such as displacements and forces, while preserving essential physical principles such as thermodynamic consistency. This work further demonstrates how data-driven approaches reveal previously unknown structure–property relationships, such as the role of dangling atomic masses in the thermal transport of crystalline nanoporous materials. Finally, it introduces a chemistry-constrained generative framework that proposes synthesizable, diverse, and novel molecular structures using limited data, while providing interpretable representations of the molecular generation.
Together, these contributions establish that physics-guided machine learning can complement and extend traditional materials science by delivering reliable, interpretable, and generalizable solutions to longstanding challenges in design and modeling.

Traditional constitutive models rely on hand-crafted parametric forms with limited expressivity and generalizability, while neural network-based models can capture complex material behavior but often lack interpretability. To balance these trade-offs, we present monotonic Input-Convex Kolmogorov-Arnold Networks (ICKANs) for learning polyconvex hyperelastic constitutive laws. ICKANs leverage the Kolmogorov-Arnold representation, decomposing the model into compositions of trainable univariate spline-based activation functions for rich expressivity. We introduce trainable monotonic input-convex splines within the KAN architecture, ensuring physically admissible polyconvex models for isotropic compressible hyperelasticity. The resulting models are both compact and interpretable, enabling explicit extraction of analytical constitutive relationships through a monotonic input-convex symbolic regression technique. Through unsupervised training on full-field strain data and limited global force measurements, ICKANs accurately capture nonlinear stress–strain behavior across diverse strain states. Finite element simulations of unseen geometries with trained ICKAN hyperelastic constitutive models confirm the framework's robustness and generalization capability.

Vitrimer is a new, exciting class of sustainable polymers with healing abilities due to their dynamic covalent adaptive networks. However, a limited choice of constituent molecules restricts their property space and potential applications. To overcome this challenge, an innovative approach coupling molecular dynamics (MD) simulations and a novel graph variational autoencoder (VAE) model for inverse design of vitrimer chemistries with desired glass transition temperature (T_g) is presented. The first diverse vitrimer dataset of one million chemistries is curated and T_g for 8,424 of them is calculated by high-throughput MD simulations calibrated by a Gaussian process model. The proposed VAE employs dual graph encoders and a latent dimension overlapping scheme which allows for individual representation of multi-component vitrimers. High accuracy and efficiency of the framework are demonstrated by discovering novel vitrimers with desirable T_g beyond the training regime. To validate the effectiveness of the framework in experiments, vitrimer chemistries are generated with a target T_g = 323 K. By incorporating chemical intuition, a novel vitrimer with T_g of 311–317 K is synthesized, experimentally demonstrating healability and flowability. The proposed framework offers an exciting tool for polymer chemists to design and synthesize novel, sustainable polymers for various applications.

Advancements in machine learning have sparked significant interest in designing mechanical metamaterials, i.e., materials that derive their properties from their inherent microstructure rather than just their constituent material. We propose a data-driven exploration of the design space of growth-based cellular metamaterials based on star-shaped distances. These two-dimensional metamaterials are based on periodically-repeating unit cells consisting of material and void patterns with non-trivial geometries. Machine learning models exploiting large datasets are then employed to inverse design growth-based metamaterials for tailored anisotropic stiffness. Firstly, a forward model is created to bypass the growth and homogenization process and accurately predict the mechanical properties given a finite set of design parameters. Secondly, an inverse model is used to invert the structure–property maps and enable the accurate prediction of designs for a given anisotropic stiffness query. We successfully demonstrate the frameworks’ generalization capabilities by inverse designing for stiffness properties chosen from outside the domain of the design space.

Within the scope of our recent approach for Efficient Unsupervised Constitutive Law Identification and Discovery (EUCLID), we propose an unsupervised Bayesian learning framework for discovery of parsimonious and interpretable constitutive laws with quantifiable uncertainties. As in deterministic EUCLID, we do not resort to stress data, but only to realistically measurable full-field displacement and global reaction force data; as opposed to calibration of an a priori assumed model, we start with a constitutive model ansatz based on a large catalog of candidate functional features; we leverage domain knowledge by including features based on existing, both physics-based and phenomenological, constitutive models. In the new Bayesian-EUCLID approach, we use a hierarchical Bayesian model with sparsity-promoting priors and Monte Carlo sampling to efficiently solve the parsimonious model selection task and discover physically consistent constitutive equations in the form of multivariate multi-modal probabilistic distributions. We demonstrate and validate the ability to accurately and efficiently recover isotropic and anisotropic hyperelastic models like the Neo-Hookean, Isihara, Gent–Thomas, Arruda–Boyce, Ogden, and Holzapfel models in both elastostatics and elastodynamics. The discovered constitutive models are reliable under both epistemic uncertainties — i.e. uncertainties on the true features of the constitutive catalog – and aleatoric uncertainties – which arise from the noise in the displacement field data, and are automatically estimated by the hierarchical Bayesian model.

We propose a new approach for unsupervised learning of hyperelastic constitutive laws with physics-consistent deep neural networks. In contrast to supervised learning, which assumes the availability of stress–strain pairs, the approach only uses realistically measurable full-field displacement and global reaction force data, thus it lies within the scope of our recent framework for Efficient Unsupervised Constitutive Law Identification and Discovery (EUCLID) and we denote it as NN-EUCLID. The absence of stress labels is compensated for by leveraging a physics-motivated loss function based on the conservation of linear momentum to guide the learning process. The constitutive model is based on input-convex neural networks, which are capable of learning a function that is convex with respect to its inputs. By employing a specially designed neural network architecture, multiple physical and thermodynamic constraints for hyperelastic constitutive laws, such as material frame indifference, material stability, and stress-free reference configuration are automatically satisfied. We demonstrate the ability of the approach to accurately learn several hidden isotropic and anisotropic hyperelastic constitutive laws – including e.g., Mooney–Rivlin, Arruda–Boyce, Ogden, and Holzapfel models – without using stress data. For anisotropic hyperelasticity, the unknown anisotropic fiber directions are automatically discovered jointly with the constitutive model. The neural network-based constitutive models show good generalization capability beyond the strain states observed during training and are readily deployable in a general finite element framework for simulating complex mechanical boundary value problems with good accuracy.