arXiv subjects

Mingjian Wen

Publications and source records attributed to Mingjian Wen.

17 recordsLinked to original sources

arXiv2026

Reusable Operators for Irreducible Cartesian Tensor Decomposition and Coupling

Molecular and material properties, from the polarizability to the elastic constants, are described by tensors. Their behavior under rotations is made explicit when a tensor is decomposed into irreducible parts that transform independently. In Cartesian form these parts are the symmetric and traceless irreducible Cartesian tensors (ICTs), whose decomposition and coupling underlie selection rules, orientational averages, and the use of symmetry. The decomposition is a textbook result at rank two, but higher rank and the intrinsic symmetry of physical tensors make it nontrivial. What has been lacking, unlike in the well-established spherical formalism, is a general construction for a given intrinsic symmetry, together with reusable operators that extract the ICTs and rebuild the original tensor exactly. Here, we develop such a construction and obtain these operators explicitly. These operators depend on rank and symmetry alone, and therefore each need only be built once and then applied to any tensor of that class. Building on them, we further obtain the Cartesian harmonics of a vector and the operators that couple two ICTs into a third, both central to equivariant machine learning. The construction is demonstrated on the elastic tensor, in both its second-order form of rank four and its third-order form of rank six. The ICTs of the rank-4 tensor also define a rotation- and scale-invariant measure of anisotropy, which we evaluate across the first-principles elastic tensors of crystalline materials from the Materials Project. The construction is implemented in the open-source package natto, which produces the operators in both exact symbolic and numerical form.

cond-mat.mtrl-sci↗

arXiv2026

Comparative study of ensemble-based uncertainty quantification methods for neural network interatomic potentials

Machine learning interatomic potentials (MLIPs) enable atomistic simulations with near first-principles accuracy at substantially reduced computational cost, making them powerful tools for large-scale materials modeling. The accuracy of MLIPs is typically validated on a held-out dataset of \emph{ab initio} energies and atomic forces. However, accuracy on these small-scale properties does not guarantee reliability for emergent, system-level behavior -- precisely the regime where atomistic simulations are most needed, but for which direct validation is often computationally prohibitive. As a practical heuristic, predictive precision -- quantified as inverse uncertainty -- is commonly used as a proxy for accuracy, but its reliability remains poorly understood, particularly for system-level predictions. In this work, we systematically assess the relationship between predictive precision and accuracy in both in-distribution (ID) and out-of-distribution (OOD) regimes, focusing on ensemble-based uncertainty quantification methods for neural network potentials, including bootstrap, dropout, random initialization, and snapshot ensembles. We use held-out cross-validation for ID assessment and calculate cold curve energies and phonon dispersion relations for OOD testing. These evaluations are performed across various carbon allotropes as representative test systems. We find that uncertainty estimates can behave counterintuitively in OOD settings, often plateauing or even decreasing as predictive errors grow. These results highlight fundamental limitations of current uncertainty quantification approaches and underscore the need for caution when using predictive precision as a stand-in for accuracy in large-scale, extrapolative applications.

cond-mat.mtrl-sci↗

arXiv2026

Guided Synthesis of EMT Zeolites by Machine Learning

Zeolites are microporous crystalline materials with diverse frameworks, widely used in industrial applications such as petroleum refining and molecular separation. Unlike most zeolites, EMT can be synthesized under mild conditions (at low temperatures and without the use of organic structure-directing agents), making it attractive for cost-effective and environmentally sustainable production. However, the specific synthesis conditions that selectively produce EMT rather than similar frameworks like FAU are not yet well established. In this work, we develop machine learning (ML) models to guide the discovery of synthesis conditions for EMT zeolites. Our dataset comprises 174 experimental synthesis attempts, recording reaction time, temperature, silica and alumina sources, Si/Al stoichiometric ratio, and other synthesis parameters. We apply both classical ML methods and pretrained foundation models to predict zeolite framework outcomes from these synthesis parameters. Feature importance analysis identifies critical parameters for EMT formation, validating known synthesis principles. Leveraging the ML models, we explore the synthesis space and identify six promising new conditions for EMT formation. Experimental validation confirms EMT crystallization in five cases, including two with Si/Al stoichiometric ratios outside the training dataset's range. Evaluation on independent literature-reported synthesis conditions further demonstrates the generalizability of the model. This work demonstrates a data-driven approach to accelerating zeolite synthesis, closing the loop between ML prediction and experimental validation.

cond-mat.mtrl-sci↗

arXiv2026

Atomistic Machine Learning with Irreducible Cartesian Natural Tensors

Atomistic machine learning is a powerful tool for accurate and efficient investigation of material behavior at the atomic scale. While attempts have been made to construct models directly within Cartesian space, they face challenges in providing a systematic framework based on irreducible representations - a core feature of widely used spherical models. Here we propose Cartesian Natural Tensor Networks to overcome these limitations and thus offer a general, symmetry-preserving framework for atomistic machine learning. We present a theory of irreducible representations using Cartesian natural tensors, comprising their construction, their products, and a systematic scheme to decompose and reconstruct high-rank physical tensors. Leveraging this machinery, we develop equivariant machine learning interatomic potentials for materials and molecular systems with performance on par with leading spherical models. It further captures accurate structure-property relationships for tensorial quantities ranging from low-rank dipole moments to high-rank tensors with complex symmetries, such as the elastic constant tensor.

cond-mat.mtrl-sci↗

arXiv2026

Strain-Dependent Ionic Transport in Li3YCl6 Solid Electrolytes

Solid-state batteries require electrolytes that sustain high ionic conductivity under the mechanical environment of a functioning cell. Lattice strain, arising from stack pressure, thermal cycling, or lattice mismatch at interfaces, can either enhance or suppress Li+ transport in solid electrolytes, yet how it couples to the underlying diffusion mechanism remains poorly understood. Using Li3YCl6 halide superionic conductor, we address this with large-scale molecular dynamics simulations driven by an Atomic Cluster Expansion (ACE) machine learning interatomic potential trained on first-principles data. The ACE model faithfully reproduces experimental and \textit{ab initio} structural, mechanical, and transport properties of Li3YCl6. We find that Li+ diffusion in Li3YCl6 follows a two-regime Arrhenius behavior, crossing over at a critical temperature $T_c$ from one-dimensional hopping at low temperature to three-dimensional cooperative diffusion at high temperature. Strain substantially modulates diffusivity: tensile strain enhances it while compressive strain suppresses it, yet leaves $T_c$ invariant, indicating that strain tunes diffusion efficiency without reshaping the underlying transport framework. In each regime, the mechanistic origin differs: altered activation barriers dominate at low temperature, while modified pre-exponential factors become critical at high temperature. These results establish lattice strain as a design lever for ionic conductivity in Li3YCl6 solid-state electrolytes.

cond-mat.mtrl-sci↗

arXiv2026

An information-matching approach to optimal experimental design and active learning

The efficacy of mathematical models heavily depends on the quality of the training data, yet collecting sufficient data is often expensive and challenging. Many modeling applications require inferring parameters only as a means to predict other quantities of interest (QoI). Because models often contain many unidentifiable (sloppy) parameters, QoIs often depend on a relatively small number of parameter combinations. Therefore, we introduce an information-matching criterion based on the Fisher Information Matrix to select the most informative training data from a candidate pool. This method ensures that the selected data contain sufficient information to learn only those parameters that are needed to constrain downstream QoIs. It is formulated as a convex optimization problem, making it scalable to large models and datasets. We demonstrate the effectiveness of this approach across various modeling problems in diverse scientific fields, including power systems and underwater acoustics. Finally, we use information-matching as a query function within an Active Learning loop for material science applications. In all these applications, we find that a relatively small set of optimal training data can provide the necessary information for achieving precise predictions. These results are encouraging for diverse future applications, particularly active learning in large machine learning models.

cs.LG↗

arXiv2025

Cartesian atomic moment machine learning interatomic potentials

Machine learning interatomic potentials (MLIPs) have substantially advanced atomistic simulations in materials science and chemistry by balancing accuracy and computational efficiency. While leading MLIPs rely on representing atomic environments using spherical tensors, Cartesian representations offer potential advantages in simplicity and efficiency. Here, we introduce the Cartesian Atomic Moment Potential (CAMP), an approach to building MLIPs entirely in Cartesian space. CAMP constructs atomic moment tensors from neighboring atoms and employs tensor products to incorporate higher body-order interactions, providing a complete description of local atomic environments. Integrated into a graph neural network (GNN) framework, CAMP enables physically motivated, systematically improvable potentials. The model demonstrates excellent performance across diverse systems, including periodic structures, small organic molecules, and two-dimensional materials, achieving accuracy, efficiency, and stability in molecular dynamics simulations that rival or surpass current leading models. CAMP provides a powerful tool for atomistic simulations to accelerate materials understanding and discovery.

cond-mat.mtrl-sci↗

arXiv2024

Fine-Tuning Language Models on Multiple Datasets for Citation Intention Classification

Citation intention Classification (CIC) tools classify citations by their intention (e.g., background, motivation) and assist readers in evaluating the contribution of scientific literature. Prior research has shown that pretrained language models (PLMs) such as SciBERT can achieve state-of-the-art performance on CIC benchmarks. PLMs are trained via self-supervision tasks on a large corpus of general text and can quickly adapt to CIC tasks via moderate fine-tuning on the corresponding dataset. Despite their advantages, PLMs can easily overfit small datasets during fine-tuning. In this paper, we propose a multi-task learning (MTL) framework that jointly fine-tunes PLMs on a dataset of primary interest together with multiple auxiliary CIC datasets to take advantage of additional supervision signals. We develop a data-driven task relation learning (TRL) method that controls the contribution of auxiliary datasets to avoid negative transfer and expensive hyper-parameter tuning. We conduct experiments on three CIC datasets and show that fine-tuning with additional datasets can improve the PLMs' generalization performance on the primary dataset. PLMs fine-tuned with our proposed framework outperform the current state-of-the-art models by 7% to 11% on small datasets while aligning with the best-performing model on a large dataset.

cs.CL↗

arXiv2024

Uncertainty Quantification and Propagation in Atomistic Machine Learning

Machine learning (ML) offers promising new approaches to tackle complex problems and has been increasingly adopted in chemical and materials sciences. Broadly speaking, ML models employ generic mathematical functions and attempt to learn essential physics and chemistry from a large amount of data. Consequently, because of the limited physical or chemical principles in the functional form, the reliability of the predictions is oftentimes not guaranteed, particularly for data far out of distribution. It is critical to quantify the uncertainty in model predictions and understand how the uncertainty propagates to downstream chemical and materials applications. Herein, we review existing uncertainty quantification (UQ) and uncertainty propagation (UP) methods for atomistic ML under a united framework of probabilistic modeling. We first categorize the UQ methods, with the aim to elucidate the similarities and differences between them. We also discuss performance metrics to evaluate the accuracy, precision, calibration, and efficiency of the UQ methods and techniques for model recalibration. With these metrics, we survey existing benchmark studies of the UQ methods using molecular and materials datasets. Furthermore, we discuss UP methods to propagate the uncertainty obtained from ML models in widely used materials and chemical simulation techniques, such as molecular dynamics and microkinetic modeling. We also provide remarks on the challenges and future opportunities of UQ and UP in atomistic ML.

cond-mat.mtrl-sci↗

arXiv2024

An equivariant graph neural network for the elasticity tensors of all seven crystal systems

The elasticity tensor that describes the elastic response of a material to external forces is among the most fundamental properties of materials. The availability of full elasticity tensors for inorganic crystalline compounds, however, is limited due to experimental and computational challenges. Here, we report the materials tensor (MatTen) model for rapid and accurate estimation of the full fourth-rank elasticity tensors of crystals. Based on equivariant graph neural networks, MatTen satisfies the two essential requirements for elasticity tensors: independence of the frame of reference and preservation of material symmetry. Consequently, it provides a unified treatment of elasticity tensors for all seven crystal systems across diverse chemical spaces, without the need to deal with each separately.. MatTen was trained on a dataset of first-principles elasticity tensors garnered by the Materials Project over the past several years (we are releasing the data herein) and has broad applications in predicting the isotropic elastic properties of polycrystalline materials, examining the anisotropic behavior of single crystals, and discovering new materials with exceptional mechanical properties. Using MatTen, we have discovered a hundred new crystals with extremely large maximum directional Young's modulus and eleven polymorphs of elemental cubic metals with unconventional spatial orientation of Young's modulus.

cond-mat.mtrl-sci↗

arXiv2022

Injecting Domain Knowledge from Empirical Interatomic Potentials to Neural Networks for Predicting Material Properties

For decades, atomistic modeling has played a crucial role in predicting the behavior of materials in numerous fields ranging from nanotechnology to drug discovery. The most accurate methods in this domain are rooted in first-principles quantum mechanical calculations such as density functional theory (DFT). Because these methods have remained computationally prohibitive, practitioners have traditionally focused on defining physically motivated closed-form expressions known as empirical interatomic potentials (EIPs) that approximately model the interactions between atoms in materials. In recent years, neural network (NN)-based potentials trained on quantum mechanical (DFT-labeled) data have emerged as a more accurate alternative to conventional EIPs. However, the generalizability of these models relies heavily on the amount of labeled training data, which is often still insufficient to generate models suitable for general-purpose applications. In this paper, we propose two generic strategies that take advantage of unlabeled training instances to inject domain knowledge from conventional EIPs to NNs in order to increase their generalizability. The first strategy, based on weakly supervised learning, trains an auxiliary classifier on EIPs and selects the best-performing EIP to generate energies to supplement the ground-truth DFT energies in training the NN. The second strategy, based on transfer learning, first pretrains the NN on a large set of easily obtainable EIP energies, and then fine-tunes it on ground-truth DFT energies. Experimental results on three benchmark datasets demonstrate that the first strategy improves baseline NN performance by 5% to 51% while the second improves baseline performance by up to 55%. Combining them further boosts performance.

cs.LG↗

arXiv2022

Extending OpenKIM with an Uncertainty Quantification Toolkit for Molecular Modeling

Atomistic simulations are an important tool in materials modeling. Interatomic potentials (IPs) are at the heart of such molecular models, and the accuracy of a model's predictions depends strongly on the choice of IP. Uncertainty quantification (UQ) is an emerging tool for assessing the reliability of atomistic simulations. The Open Knowledgebase of Interatomic Models (OpenKIM) is a cyberinfrastructure project whose goal is to collect and standardize the study of IPs to enable transparent, reproducible research. Part of the OpenKIM framework is the Python package, KIM-based Learning-Integrated Fitting Framework (KLIFF), that provides tools for fitting parameters in an IP to data. This paper introduces a UQ toolbox extension to KLIFF. We focus on two sources of uncertainty: variations in parameters and inadequacy of the functional form of the IP. Our implementation uses parallel-tempered Markov chain Monte Carlo (PTMCMC), adjusting the sampling temperature to estimate the uncertainty due to the functional form of the IP. We demonstrate on a Stillinger--Weber potential that makes predictions for the atomic energies and forces for silicon in a diamond configuration. Finally, we highlight some potential subtleties in applying and using these tools with recommendations for practitioners and IP developers.

physics.comp-ph↗

arXiv2022

Bayesian, frequentist, and information geometric approaches to parametric uncertainty quantification of classical empirical interatomic potentials

In this paper, we consider the problem of quantifying parametric uncertainty in classical empirical interatomic potentials (IPs) using both Bayesian (Markov Chain Monte Carlo) and frequentist (profile likelihood) methods. We interface these tools with the Open Knowledgebase of Interatomic Models and study three models based on the Lennard-Jones, Morse, and Stillinger--Weber potentials. We confirm that IPs are typically sloppy, i.e., insensitive to coordinated changes in some parameter combinations. Because the inverse problem in such models is ill-conditioned, parameters are unidentifiable. This presents challenges for traditional statistical methods, as we demonstrate and interpret within both Bayesian and frequentist frameworks. We use information geometry to illuminate the underlying cause of this phenomenon and show that IPs have global properties similar to those of sloppy models from fields such as systems biology, power systems, and critical phenomena. IPs correspond to bounded manifolds with a hierarchy of widths, leading to low effective dimensionality in the model. We show how information geometry can motivate new, natural parameterizations that improve the stability and interpretation of uncertainty quantification analysis and further suggest simplified, less-sloppy models.

cond-mat.mtrl-sci↗

arXiv2021

KLIFF: A framework to develop physics-based and machine learning interatomic potentials

Interatomic potentials (IPs) are reduced-order models for calculating the potential energy of a system of atoms given their positions in space and species. IPs treat atoms as classical particles without explicitly modeling electrons and thus are computationally far less expensive than first-principles methods, enabling molecular simulations of significantly larger systems over longer times. Developing an IP is a complex iterative process involving multiple steps: assembling a training set, designing a functional form, optimizing the function parameters, testing model quality, and deployment to molecular simulation packages. This paper introduces the KIM-based learning-integrated fitting framework (KLIFF), a package that facilitates the entire IP development process. KLIFF supports both physics-based and machine learning IPs. It adopts a modular approach whereby various components in the fitting process, such as atomic environment descriptors, functional forms, loss functions, optimizers, quality analyzers, and so on, work seamlessly with each other. This provides a flexible framework for the rapid design of new IP forms. Trained IPs are compatible with the Knowledgebase of Interatomic Models (KIM) application programming interface (API) and can be readily used in major materials simulation packages compatible with KIM, including ASE, DL_POLY, GULP, LAMMPS, and QC. KLIFF is written in Python with computationally intensive components implemented in C++. It is parallelized over data and supports both shared-memory multicore desktop machines and high-performance distributed memory computing clusters. We demonstrate the use of KLIFF by fitting a physics-based Stillinger--Weber potential and a machine learning neural network potential for silicon. The KLIFF package, together with its documentation, is publicly available at: https://github.com/openkim/kliff.

cond-mat.mtrl-sci↗

arXiv2019

Hybrid neural network potential for multilayer graphene

Monolayer and multilayer graphene are promising materials for applications such as electronic devices, sensors, energy generation and storage, and medicine. In order to perform large-scale atomistic simulations of the mechanical and thermal behavior of graphene-based devices, accurate interatomic potentials are required. Here, we present a new interatomic potential for multilayer graphene structures referred to as "hNN--Gr$_x$." This hybrid potential employs a neural network to describe short-range interactions and a theoretically-motivated analytical term to model long-range dispersion. The potential is trained against a large dataset of monolayer graphene, bilayer graphene, and graphite configurations obtained from ab initio total-energy calculations based on density functional theory (DFT). The potential provides accurate energy and forces for both intralayer and interlayer interactions, correctly reproducing DFT results for structural, energetic, and elastic properties such as the equilibrium layer spacing, interlayer binding energy, elastic moduli, and phonon dispersions to which it was not fit. The potential is used to study the effect of vacancies on thermal conductivity in monolayer graphene and interlayer friction in bilayer graphene. The potential is available through the OpenKIM interatomic potential repository at \url{https://openkim.org}.

cond-mat.mtrl-sci↗

arXiv2018

Dihedral-angle-corrected registry-dependent interlayer potential for multilayer graphene structures

The structural relaxation of multilayer graphene is essential in describing the interesting electronic properties induced by intentional misalignment of successive layers, including the recently reported superconductivity in twisted bilayer graphene. This is difficult to accomplish without an accurate interatomic potential. Here, we present a new, registry-dependent Kolmogorov-Crespi type interatomic potential to model interlayer interactions in multilayer graphene structures. It consists of two parts representing attractive interaction due to dispersion, and repulsive interaction due to anisotropic overlap of electronic orbitals. An important new feature is a dihedral-angle-dependent term that is added to the repulsive part in order to describe correctly several distinct stacking states that the original Kolmogorov-Crespi potential cannot distinguish. We refer to the new model as the Dihedral-angle-corrected Registry-dependent Interlayer Potential (DRIP). Computations for several test problems show that DRIP correctly reproduces the binding, sliding, and twisting energies and forces obtained from ab initio total-energy calculations based on density functional theory. We use the new potential to study the structural properties of a twisted graphene bilayer and the exfoliation of graphene from graphite. Our potential is available through the OpenKIM interatomic potential repository at https://openkim.org.

cond-mat.mtrl-sci↗

arXiv2016

A KIM-compliant potfit for fitting sloppy interatomic potentials: Application to the EDIP model for silicon

Fitted interatomic potentials are widely used in atomistic simulations thanks to their ability to compute the energy and forces on atoms quickly. However, the simulation results crucially depend on the quality of the potential being used. Force matching is a method aimed at constructing reliable and transferable interatomic potentials by matching the forces computed by the potential as closely as possible, with those obtained from first principles calculations. The potfit program is an implementation of the force-matching method that optimizes the potential parameters using a global minimization algorithm followed by a local minimization polish. We extended potfit in two ways. First, we adapted the code to be compliant with the KIM Application Programming Interface (API) standard (part of the Knowledgebase of Interatomic Models Project). This makes it possible to use potfit to fit many KIM potential models, not just those prebuilt into the potfit code. Second, we incorporated the geodesic Levenberg--Marquardt (LM) minimization algorithm into potfit as a new local minimization algorithm. The extended potfit was tested by generating a training set using the KIM Environment-Dependent Interatomic Potential (EDIP) model for silicon and using potfit to recover the potential parameters from different initial guesses. The results show that EDIP is a "sloppy model" in the sense that its predictions are insensitive to some of its parameters, which makes fitting more difficult. We find that the geodesic LM algorithm is particularly efficient for this case. The extended potfit code is the first step in developing a KIM-based fitting framework for interatomic potentials for bulk and two dimensional materials. The code is available for download via https://www.potfit.net.

cond-mat.other↗