Publications

We develop a physics-informed Monte Carlo Tree Search (MCTS) framework for inverse design of correlated materials. The method uses the MACE-MP-0 machine learning potential as an efficient surrogate for density functional theory (DFT), enabling rapid evaluation of formation energies and electronic properties during search. The algorithm employs multi-objective reward functions incorporating formation energy, convex hull stability, and density of states characteristics near the Fermi level. Ensemble scoring across diverse weight configurations ensures recommendation robustness. We first validate the framework on a uranium-constrained design space (108 compounds), where MCTS identified all globally optimal compositions while evaluating only 65% of candidates. Experimental synthesis validated predictive capability, with successful growth of four single-crystal compounds (UV6Sn6, UNb6Sn6, UCr6Ge6, UCo6Ge6). Extension to the full lanthanide-U space (1,728 compounds) demonstrated greater efficiency: MCTS explored only 37% of possible chemistries while recovering Pareto-optimal solutions and identifying chemically diverse high-performing candidates. Equal-weight optimization identified LuIr6Si6 as the top-ranked compound (composite score 0.920), exhibiting exceptional formation energy, near-hull stability, and high DOS reward. Comparison with literature-synthesized compounds confirmed that MCTS recommendations occupy thermodynamically comparable regimes to known materials, though experimental accessibility depends on additional kinetic and synthesis factors beyond computational reward functions. This work establishes intelligent tree search algorithms as an effective strategy for quantum materials discovery, offering efficiency gains over high-throughput enumeration while enabling exploration guided by physics-based priors and learned structure-property relationships.

Machine-learned interatomic potentials (MLIPs) deliver near-quantum-mechanical accuracy at a fraction of ab initio computational cost. Their performance depends on the quality and composition of the training data: the training set must cover the range of atomic environments the model will encounter, yet unnecessary redundancy inflates training cost without improving accuracy. Methods based on information-theoretic diversity measures require specification of a length-scale parameter (kernel bandwidth) that controls how similar two atomic environments must be before one is considered redundant; the optimal value varies across structurally distinct phases, making automated pruning of heterogeneous datasets challenging. We present a pruning framework that adapts the bandwidth independently for each thermodynamic group (phase, density, temperature) using an automated elbow criterion applied to QUESTS information-entropy metrics, then applies farthest-point sampling within each group to select maximally diverse atomic environments. Using the Small Carbon ChIMES 2.0 dataset, we show a single global bandwidth overestimates configuration uniqueness and preferentially retains liquid environments at the expense of crystalline phase coverage. Relative to global pruning, group-aware adaptive pruning reduces test force-MAE by up to 37% for diamond and 26% for graphite when training on only 5% of the original data, with corresponding improvements in radial distribution function fidelity.

Machine-learned interatomic potentials (ML-IAPs) have emerged as a powerful tool for achieving nominally quantum-accurate simulations at reduced computational cost. However, for covalently bonded systems, a fundamental tension exists between the model size required to resolve highly featured short-range interactions and the extended interaction range needed to capture smoother long-range contributions. This work introduces a multi-layer representation that resolves this tension by decomposing the potential energy surface into two overlaid models: a short-range layer employing a dense basis to capture bond rearrangements and repulsion, and a long-range layer employing a sparser basis for smoothly varying contributions. This strategy is implemented within the ChIMES ML-IAP framework and demonstrated on three systems of increasing complexity: a classical united-atom propane model, water across non-reactive and reactive thermodynamic conditions, and reactive C/O mixtures spanning a broad range of temperatures, pressures, and compositions. In each case, multi-layer models achieve accuracy comparable to single-layer models while yielding at least an order-of-magnitude reduction in computational cost. A comprehensive hyperparameter sensitivity study on the propane system provides physically motivated heuristics for selecting additional hyperparameters introduced through the multi-layer strategy. Taken together, these results establish multi-layer ChIMES as a general and practical strategy for improving the efficiency of bespoke ML-IAPs without sacrificing predictive accuracy.

While significant strides have been made to develop universal (foundation) interatomic potentials, these models are often too computationally expensive for large-scale molecular dynamics or lack the necessary accuracy for certain applications due to the use of overly-general training sets. We develop a multi-level optimization algorithm that uses cluster-based training to efficiently learn training data weights that are optimally informative for the properties of interest while simultaneously optimizing model hyperparameters and maintaining transferability. Demonstrated on a bespoke SNAP model for carbon, this approach shows an order-of-magnitude improvement in prediction of force constants and radial distribution functions over a baseline model trained on the full dataset. We compare HDBSCAN and k-means clustering with bispectrum descriptors, autoencoder embeddings, and pre-trained interatomic potential embeddings, finding HDBSCAN performs best with structure descriptors that closely match those used during inference.

Machine-learned interatomic models represent a significant advancement in simulation methods, extending the predictive ability of first-principles methods to previously inaccessible length and time scales. However, the data-driven nature of these models can lead to difficult-to-detect errors that compromise prediction accuracy. We introduce a novel fingerprinting approach based on the ChIMES ML-IAM graph-based descriptor that enables efficient, statistically rigorous analysis of configurations used in ML-IAM training and those generated in application (e.g., molecular dynamics simulations). We show these fingerprints effectively assess the novelty of a configuration relative to an existing dataset and quantify dissimilarity among configurations — two key tasks for active learning, dataset curation, and on-the-fly uncertainty quantification.

We explore the use of tools from Riemannian geometry for the analysis of symmetric positive definite (SPD) matrices. An SPD matrix is a versatile data representation commonly used in chemical engineering (e.g., covariance/correlation/Hessian matrices and images), and powerful techniques are available for its analysis (e.g., principal component analysis). A key observation motivating this work is that SPD matrices live on a Riemannian manifold, and exploiting this basic property can yield significant benefits in data-centric tasks such as classification and dimensionality reduction. We demonstrate this via case studies that conduct anomaly detection in the context of process monitoring and image analysis.