Session K60: Machine Learning of Molecules and Materials: Materials I

Overcoming the limits of approximate electronic structure models in machine learning accelerated materials discovery

Machine learning (ML)-accelerated discovery of transition metal containing materials such as light-harvesting chromophores, phosphors, and other photoactive complexes holds great promise. Nevertheless, the open shell d electrons that impart many of the desirable properties to these systems also make them notoriously challenging to study with conventional electronic structure techniques such as density functional theory (DFT). Thus, when ML is used to accelerate computational screening, it often inherits the biases of the underlying method used to generate training data. I will describe our recent efforts to overcome these limits through three complementary approaches. First, I will describe how we have developed machine learning models trained directly on experimental reports of iridium phosphors, leading to the development of ML models that can predict experimental emission energies and lifetimes with superior or equivalent performance to conventional methods such as time-dependent DFT but in a fraction of the computational time. Next, I will describe how we overcome limits of DFT uncertainty in screening for light harvesting chromophores with earth abundant 3d metals by incorporating method insensitivity into a multi-objective optimization workflow, requiring a consensus of functionals to agree on a property in order for a material to be selected as optimal. These workflows accelerate materials discovery by at least 1000-fold. Finally, I will describe our development of a density functional "recommender" that can identify which DFT functional is most predictive to obtain accurate vertical spin excitation energies in transition metal complexes.    

Accelerating Computational Chemistry and Materials Science Research with Azure Quantum Elements

This talk introduces Azure Quantum Elements (AQE), a new system for computational chemistry and materials science that combines the power of natural language AI assistance, artificial intelligence, high-performance computing, and quantum-classical workflows to accelerate scientific discovery. Materials simulations often require screening millions of candidate molecules and performing billions of force/energy evaluations per molecule. AQE achieves scale by providing access to cloud-optimized code and state-of-the-art automation. Speedups are achieved with AI methods that replace expensive quantum chemistry calculations of simulation forces and energies with much faster AI models. AQE also provides tools to understand which chemistry and materials problems require quantum computers, and enables exploration of quantum problems on noisy small-scale quantum devices. The audience will also learn about recent uses of AQE to expand research and development capabilities in areas such as designing new chemical systems to reduce carbon emissions, accelerating drug discovery, and speeding up hydrogen fuel cell innovation.   

Accelerating materials discovery using integrated deep machine learning approaches

We present an integrated deep machine learning (ML) approach that combines crystal graph convolutional neural networks (CGCNN) for predicting formation energies and artificial neural networks (ANN) for constructing interatomic potentials. Using the La–Si–P ternary system as a proof-of-concept, we achieve a remarkable speed-up of at least 100 times compared to high-throughput first-principles calculations. The ML approach successfully identifies known compounds and uncovers 16 new P-rich compounds with formation energies within 100 meV per atom above the convex hull, including a stable La₂SiP₃ phase. We also employ the developed ML interatomic potential in a genetic algorithm for efficient structure search, leading to the discovery of more metastable compounds. Moreover, substitution of La atoms with Ba reveals a new stable quaternary compound, BaLaSiP₃. Our generic and robust approach holds great promise for accelerating materials discovery in various compounds, enabling more efficient exploration of complex chemical spaces and enhancing the prediction of material properties.   

Equivariant Graph Neural Networks for Predicting Spin-Crossover Energy in Transition Metal Complexes

The behavior of spin-crossover materials is influenced by the choice of ligands coordinating a metallic center. Thus, spin-crossover can be tailored chemically to a wide range of application interests [J. Matter. Chem. C 2, 1360 (2014)]. However, the experimental exploration of such a vast combinatorial chemical space is cumbersome and expensive. Machine learning (ML) algorithms offer an alternative that can characterize the relationships between ligands and a property of interest [J. Phys. Chem. Lett. 9, 1064 (2018), Ind. Eng. Chem. Res. 57, 13973 (2018), Chem. Rev. 121, 9927 (2021)]. We describe the implementation of a two-step ML approach based on equivariant graph networks [arXiv:2102.09844v3]. First, an autoencoder is trained on a vast materials dataset. Then, we use the encoded embedding space to predict the adiabatic spin-crossover energy using a different dataset of fifteen hundred metal-organic solids, computed with the r²SCAN meta-generalized gradient density functional approximation [J. Phys. Chem. Lett. 11, 8208 (2020)]. Our efforts elucidate the efficacy of equivariant graph networks to predict novel spin crossover materials. This research has the potential to open new avenues for the efficient exploration of complex chemical spaces.   

Incorporating explicit electrostatic interactions in machine learning potentials

Long-range interactions such as electrostatics have long been a concern in developing accurate, efficient machine learning potential energy surfaces (ML-PES). Many approaches for incorporating such interactions into the structure of an ML-PES have been proposed over the past decade; however, no approach has yet exhibited the combination of accuracy, generality, and conceptual simplicity necessary to find wide acceptance. In this work, we revisit one of the earliest and simplest approaches, namely, machine learning local parameters (charges) that are incorporated into a simple functional form [1], which has recently found success in the context of van der Waals interactions [2]. We test this approach on lithium-intercalated graphite, a model system for battery electrodes, where experimental data is widely available, and explore the impact of electrostatic interactions on both the dimensional changes and the Li filling pattern. Finally, we discuss wider implications for incorporating long-range interactions in future machine learning models.

[1] N. Artrith, T. Morawietz, and J. Behler, Physical Review B 83, 153101 (2011).

[2] H. Muhli, X. Chen, A. P. Bartók, P. Hernández-León, G. Csányi, T. Ala-Nissila, and M. A. Caro, Phys. Rev. B 104, 054106 (2021).   

Designing Coarse-Grained Representations for Soft Materials using Attentive Message-Passing

Bottom-up methods for coarse-grained (CG) molecular modeling are needed to establish rigorous links between atomistic reference data and reduced molecular representations. For a given molecule, the ideal CG representation is a function of both the conformational ensemble of the system and the target physical observable(s) to be reproduced at the CG resolution, and serves as the information bottleneck limiting the precision of determining structure-function relationships. However, there is an absence of algorithms for selecting CG representations of molecules from which complex properties, including molecular electronic structure, can be accurately modeled. We describe how Graph Neural Network (GNN) architectures, which are already commonly used for molecular property prediction, can be modified using attention functions coupled to the loss of information associated with the coarsening of a representation. These methods can be used to perform arbitrary property predictions while regressing on automatically-identified CG representations or inferring representations over diverse chemistries. Sampled attention embeddings provide an interpretable framework for quantifying the space of effective representations and the collective variables influencing molecular properties.   

ML Gradients in Molecular Simulations

The success of deep learning is predicated on differentiable programming and gradient-based optimization. In scientific applications, merging machine learning models and physics-based simulators is particularly compelling. ML surrogates can replace expensive simulators and physics-derived concepts and invariances add inductive bias to otherwise black-box models. Here, we will describe research examples of exploiting ML surrogate functions, and in particular their gradients, accessed through differentiable programming in molecular simulations. Applications include active learning of machine learning potentials for ground and excited states with differentiable uncertainty, and learning of data-driven collective variables for enhanced sampling simulations.   

Transferable diversity – a data-driven representation of chemical space

While transferability in general chemistry machine learning should benefit from diverse training data, a rigorous understanding of transferability together with its interplay with chemical representation remains an open problem. In this talk, I will introduce a transferability framework and apply it to a controllable data-driven model for developing density functional approximations (DFAs) [1]. This framework reveals that human intuition introduces chemical biases that can hamper the transferability of data-driven DFAs, and it allowed us to identify strategies for their elimination. I will also show that the uncritical use of large training sets can actually hinder the transferability of DFAs, in contradiction to typical "more is more" expectations. Finally, I will demonstrate how our transferability framework yields transferable diversity, a cornerstone principle for data curation for developing general-purpose machine learning models in chemistry.   

Active-Learning for Machine-Learned Interatomic Potentials; The Example of Strongly Anharmonic Materials

Machine-learned interatomic potentials (MLIP) promise to perform efficient molecular dynamics simulations with the accuracy of ab initio methods for large supercells and long time spans, which are not feasible with ab initio methods. For strongly anharmonic materials, it is crucial to capture rare anharmonic effects, such as the formation of intrinsic defects and dynamical precursors to phase transitions [1]. To expedite the training process for such rare events, we developed an active-learning scheme that utilizes molecular dynamics with MLIP (MLIP-MD) and incorporates a measure of uncertainty to identify qualitative deviations from the model's trained region. Our applications to KCaF3 and CuI demonstrate a favorable training speed. Our MLIP-MD runs adeptly and captures essential dynamical features, including anharmonicity measure for molecular dynamics trajectory, and anharmonic vibrations, as well as rare anharmonic events. We also discuss how to employ this approach for predicting the electrical conductivity of strongly anharmonic materials using the ab initio Kubo-Greenwood approach.

[1] F. Knoop et al. Phys. Rev. Lett. 130, 236301 (2023).   

Electronic Structures of Ternary Compounds GeSbTe Based on Machine Learning Empirical Pseudopotentials

Germanium-antimony-telluride (GST) compounds have long been recognized as one of candidate materials for nonvolatile phase-change memory due to high read and write speed and low power consumption. In this study, we present electronic structure calculations of ternary GST compounds using a machine learning empirical pseudopotential method (ML-EPM) [Kim and Son, arXiv:2306.04426 (2023)]. The newly developed ML-EPM method overcomes poor transferability of traditional EPM by ML while retaining its merit such as formal simplicity and less demanding resources. We extend a previous use of ML-EPM from binary to ternary compounds. With a training set of ab initio electronic structures of various GST compounds and their rotation-covariant descriptors, we successfully generate versatile and transferable empirical pseudopotentials of Ge, Sb and Te, respectively. We demonstrate that, using the ML-EPM, computed electronic energy bands and wavefunctions of unlearned GST compounds without cumbersome self-consistency show good agreements with results from first-principles calculations. This agreement holds even for GST crystal structures with distinctive local atomic environments or more extended systems compared to those in training dataset.   

Anharmonicity in cubic boron arsenide: a machine-learning based force-field study

Due to a thermal conductivity that rivals diamond, cubic boron arsenide (c-BAs) has the potential to compete in the search for next-gen semi-conducting materials, where proper heat-management has become an increasingly important selection criterion. Based on experiments and theoretical predictions using the Boltzmann-transport equation (BTE) the high thermal conductivity is the result of a fortunate combination of effects: while c-BAs is in principle strongly anharmonic, as suggested by the temperature dependence of the thermal conductivity, the resulting three- and four-phonon scattering is suppressed effectively due to bunching of the acoustic modes and a large frequency gap between the acoustic and optical modes [1]. Alternatively, the lattice dynamics, vibrational properties and thermal conductivity can be obtained from molecular dynamics, since the required information is encoded in the trajectory of the atoms.

In this talk we are going to present our machine-learning based force-field calculations for the dynamics and thermal transport of c-BAs. The forcefield is based on the well-known lattice dynamics expansion of the total energy. With this particular choice we can obtain the higher-order force-constant tensors directly and gain further insight into the role of anharmonicity in c-BAs.

[1] L. Lindsay et al., Phys. Rev. Lett. 111, 025901 (2013)