TL;DR
This work pursues a multi-tier development strategy in which machine learning (ML) algorithms are combined with exact physical symmetries and constraints to significantly accelerate computations of electronic structure and atomistic dynamics.
It is argued that developing thermodynamics-informed ML constitutes one of the most important and least explored frontiers in materials discovery and that the next generation of ML models must move beyond static energy predictions towards a thermodynamic description of materials behaviour under realistic operating conditions.
Machine learning force fields (MLFFs) combine the high accuracy of first-principles methods with the high efficiency of classical force fields, offering new opportunities for atomic-level studies of inorganic crystalline materials. We systematically summarize the research progress on MLFFs, elucidate their fundamental principles and developmental history, and categorically introduce the technical characteristics of representative models and relevant benchmarking platforms. We aim to review the advantages of MLFFs in overcoming traditional computational limitations across four domains: structural prediction and optimization, physical properties, defect and interface properties, and phase transitions and kinetic processes. The challenges of MLFFs are also examined in computational efficiency and simulation scale, accuracy and generalization ability, data requirements and training samples, model interpretability, and physical constraints, which offer a reference for the research and application of MLFFs in the field of inorganic crystalline materials.
The limits of equivariant MLIPs are examined, and a family of foundation potentials in the NequIP and Allegro equivariant MLIP architectures are presented which achieve leading inference speeds and strong scalability as well as excellent accuracies across a range of community benchmarks.
A machine learning approach is presented that accelerates DFTB simulations by predicting optimal initial atomic charges and demonstrates that ML-predicted initial charges consistently and significantly improve SCC convergence across diverse chemical systems including organic molecules, biomolecules, water clusters, transition metal oxides and solid electrolytes.
Molecular crystals are bound by dispersion (van der Waals) interactions, whose weak nature gives rise to polymorphism, the ability of a compound to crystallize in different structures. Crystal structure profoundly influences the physical and chemical properties, and hence the functionality of molecular solids in applications including pharmaceuticals, electronic devices, and energetic materials. Therefore, the ability to predict the structure and properties of molecular crystals is of paramount importance. To this end, we combine first principles simulations with machine learning. Molecular crystal structure prediction (CSP) is challenging because it requires searching a high-dimensional configuration space with high accuracy. CSP workflow have two main components, structure generation and stability ranking. For structure generation, we develop the Genarris code [1], which generates random structures in all compatible space groups with physical constraints on intermolecular distances. Machine learned interatomic potentials (MLIPs) are trained on large data sets of first principles simulations [2], typically density functional theory (DFT) to achieve DFT-level accuracy at the computational cost of classical force fields. We have interfaced Genarris with several types of MLIPs for geometry optimization and stability ranking [1,3,4]. We have shown that MLIPs can completely replace both early-stage screening with classical force fields and final ranking with DFT, paving the way to high-throughput CSP [4]. One of the optoelectronic applications of molecular crystals is singlet fission (SF), the conversion of one photogenerated singlet exciton into two triplet excitons. SF has the potential to increase the efficiency of solar cells by harvesting two charge carriers from one high-energy photon, whose excess energy would otherwise be lost to heat. The realization of SF-based solar cells is hindered by the dearth of suitable materials. The excited-state properties of molecular crystals can be calculated using many-body perturbation theory (MBPT) within in the GW approximation and the Bethe-Salpeter equation (BSE) [5]. The computational cost of GW+BSE is prohibitive for large-scale exploration of the chemical space, and also for generating large amounts of training data. This calls for ML approaches that work well with small data.
A notably simple procedure, a method the authors refer to as deletions, yields superior performance over an array of alternative extraction methods for extracting atomic environments from large, bulk configurations and embedding them into smaller configurations suitable for DFT calculations with periodic boundary conditions.