Machine learning interatomic potentials (ML-IAPs) are emerging as transformative tools in materials modeling, promising quantum-level accuracy at a fraction of the computational cost. However, their ability to generalize beyond equilibrium configurations and to reliably capture defect- and temperature-driven behavior remains underexplored. Here, we develop and benchmark two state-of-the-art ML-IAPs, Spectral Neighbor Analysis Potential (SNAP) and Allegro, on a comprehensive dataset for monolayer MoSe2. Using density functional theory (DFT) as the reference, we evaluate their performance in capturing stress-strain behavior, phase transition energetics, defect evolution, edge stability, and fracture toughness. Allegro, a deep equivariant neural network potential, surpasses both SNAP and the classical Tersoff potential in accuracy, efficiency, and transferability. Importantly, both ML potentials accurately reproduce experimental fracture measurements and ab initio predictions of inversion domain formation-phenomena well beyond their training sets. Our findings establish ML-IAPs as viable replacements for traditional force fields in the study of non-equilibrium mechanical phenomena, enabling large-scale, high-fidelity simulations in 2D materials and beyond. This work provides a broadly applicable framework for data-driven modeling of structural and functional transformations under extreme conditions.

New materials and molecules with improved properties are needed to continue advancing society in such fields as energy, space exploration, and electronics. Unfortunately, the typical process of discovery and characterization of new material is time-consuming, which has hindered the co-design of material and product or material and manufacturing. Predictive models using machine learning (ML) capable of forecasting the properties of new materials before they are synthesized have the potential to accelerate this process and enable co-design. However, the introduction of new materials is invariably a multi-objective process and needs to balance performance with sustainability metrics, including economic, environmental, and human health. Most ML efforts to accelerate discovery have focused on performance. This paper aggregates databases of toxicology, global warming potential (GWP), and fundamental physicochemical properties of molecules and uses the combined resource to train multi-task ML models. The models are trained for 617 classification and 30 regression metrics and outperform most current single-property models. Also, a chemical extraction study was performed using per- and polyfluoroalkyl substances to investigate the limits of our model when presented with molecules that are outside of the chemical space contained within our training data. We believe we have demonstrated that the data aggregation and models can accelerate the introduction of new materials in a wide range of applications.

Melt processing of molecular crystals has several advantages over alternative routes for manufacturing materials such as pharmaceuticals, organic photovoltaics, and energetic materials. The experimental characterization of the materials properties required to assess melt processability (melting temperature, boiling temperature, decomposition temperature and vapor pressure) for the 1.3 million known molecular crystals is unfeasible; in fact, our survey of the research literature and open databases resulted in only 43 molecular materials with experimentally measured properties that satisfy a common criterion for melt-casting. We developed multi-task, graph-based neural network models that simultaneously predict these properties using a molecular graph as the only input. Screening databases of known molecules with our ML model resulted 2532 melt-castable candidates, with melting temperature between 343 K and 393 K, boiling and decomposition temperature greater than 453 K, and vapor pressure less than 0.0005 mmHg. Going beyond the space of known molecules, we apply our model with a generative approach to the CHNO chemical space, we discover 55 745 additional novel candidates with promising melt-castable characteristics. This three-orders-of-magnitude expansion highlights the power of coupling ML screening and generative design to accelerate materials discovery.

Density functional theory (DFT) is widely used to study the behavior of materials at high pressures, complementing challenging and often costly experiments. While the accuracy of DFT and the effect of various approximations and corrections have been extensively studied for materials properties around ambient conditions, few studies quantified accuracy at high pressures. We focus on the accuracy of predicted equations of state (EOS) of selected materials up to the hundred GPa regime and the description of pressure-induced phase transformations. We characterize the effect of exchange-correlation functionals, pseudopotentials, dispersion and Hubbard U correction and find that lessons-learned at ambient conditions do not always translate into the high-pressure regime. We find that the Perdew-Burke-Erzerhof solid version of the generalized gradient approximation (GGA) yields the best performance in both EOS and transformation pressure compared to Perdew-Burke-Erzerhof version of GGA, local density approximations (LDA), and the Heyd-Scuseria-Ernzerhof (HSE) hybrid functional. Adding dispersion corrections known as D2 and D3 does not improve the results. Interestingly, the local density approximation performed remarkably well. We also find that the Hubbard-U correction as a significant effect on transformation pressures in strongly correlated materials systems, indicating that the U parameter must be chosen carefully. An important by-product of this study is a FAIR repository of high-pressure simulations database on nanoHUB.

Accurately predicting the non-equilibrium mechanical properties of two-dimensional (2D) materials is essential for understanding their deformation, thermo-mechanical properties, and failure mechanisms. In this study, we parameterize and evaluate two machine learning (ML) interatomic potentials, SNAP and Allegro, for modeling the non-equilibrium behavior of monolayer MoSe2. Using a density functional theory (DFT) derived dataset, we systematically compare their accuracy and transferability against the physics-based Tersoff force field. Our results show that SNAP and Allegro significantly outperform Tersoff, achieving near-DFT accuracy while maintaining computational efficiency. Allegro surpasses SNAP in both accuracy and efficiency due to its advanced neural network architecture. Both ML potentials demonstrate strong transferability, accurately predicting out-of-sample properties such as surface stability, inversion domain formation, and fracture toughness. Unlike Tersoff, SNAP and Allegro reliably model temperature-dependent edge stabilities and phase transformation pathways, aligning closely with DFT benchmarks. Notably, their fracture toughness predictions closely match experimental measurements, reinforcing their suitability for large-scale simulations of mechanical failure in 2D materials. This study establishes ML-based force fields as a powerful alternative to traditional potentials for modeling non-equilibrium mechanical properties in 2D materials.

Pressure-induced phase transformations in materials are of interest in a range of fields, including geophysics, planetary sciences, and shock physics. In addition, the high-pressure phases can exhibit desirable properties, eliciting interest in materials science. Despite its importance, the process of finding new high-pressure phases, either experimentally or computationally, is time-consuming and often driven by intuition. In this study, we use graph neural networks trained on density functional theory (DFT) equation of state data of 2258 materials and 7255 phases to identify potential phase transitions. The model is used to explore possible phase transitions in 7677 pairs of phases and promising cases are confirmed or denied via DFT calculations. Importantly, the new data is added to the training set, the model is refined, and a new cycle of discovery is started. Within 13 iterations, we discovered 28 new high-pressure stable phases (never synthesized through high-pressure routes nor reported in high-pressure computational works) and rediscovered 18 pressure-induced phase transitions. The results provide new insight and classification of pressure-induced phase transitions in terms of the ambient properties of the phases involved.

Data-driven approaches such as deep learning can result in predictive models for material properties with exceptional accuracy and efficiency. However, in many applications, data is sparse, severely limiting their accuracy and applicability. To improve predictions, techniques such as transfer learning and multitask learning have been used. The performance of multitask learning models depends on the strength of the underlying correlations between tasks and the completeness of the data set. Standard multitask models tend to underperform when trained on sparse data sets with weakly correlated properties. To address this gap, we fuse deep-learned embeddings generated by independent pretrained single-task models, resulting in a multitask model that inherits rich, property-specific representations. By reusing (rather than retraining) these embeddings, the resulting fused model outperforms standard multitask models and can be extended with fewer trainable parameters. We demonstrate this technique on a widely used benchmark data set of quantum chemistry data for small molecules as well as a newly compiled sparse data set of experimental data collected from literature and our own quantum chemistry and thermochemical calculations.

Predictive models for the performance of explosives and propellants are important for their design, optimization, and safety. Thermochemical codes can predict some of these properties from fundamental quantities such as density and formation energies that can be obtained from first principles. Models that are simpler to evaluate are desirable for efficient, rapid screening of material screening. In addition, interpretable models can provide insight into the physics and chemistry of these materials that could be useful to direct new synthesis. Current state-of-the-art performance models are based on either the parametrization of physics-based expressions or data-driven approaches with minimal interpretability. We use parsimonious neural networks (PNNs) to discover interpretable models for the specific impulse of propellants and detonation velocity and pressure for explosives using data collected from the open literature. A combination of evolutionary optimization with custom neural networks explores and trains models with objective functions that balance accuracy and complexity. For all three properties of interest, we find interpretable models that are Pareto optimal in the accuracy and simplicity space.

Data science and artificial intelligence are playing an increasingly important role in the physical sciences. Unfortunately, in the field of energetic materials data scarcity limits the accuracy and even applicability of ML tools. To address data limitations, we compiled multi-modal data: both experimental and computational results for several properties. We find that multi-task neural networks can learn from multi-modal data and outperform single-task models trained for specific properties. As expected, the improvement is more significant for data-scarce properties. These models are trained using descriptors built from simple molecular information and can be readily applied for large-scale materials screening to explore multiple properties simultaneously. This approach is widely applicable to fields outside energetic materials.

Ge-Sb-Te (GST) alloys are leading phase-change materials for data storage due to the fast phase transition between amorphous and crystalline states. Ongoing research aims at improving the stability of the amorphous phase to improve retention. This can be accomplished by the introduction of carbon as a dopant to Ge2Sb2Te5, which is known to alter the short- and mid-range structure of the amorphous phase and form covalently bonded C clusters, both of which hinder crystallization. The relative importance of these processes as a function of C concentration is not known. We used molecular dynamics simulation based on density functional theory to study how carbon doping affects the atomic structure of GST-C. Carbon doping results in an increase in tetrahedral coordination, especially of Ge atoms, and this is known to stabilize the amorphous phase. We observe an unexpected, non-monotonous trend in the number of tetrahedral bonded Ge with the amount of carbon doping. Our simulations show an increase in the number of tetrahedral bonded Ge up to 5 at.% C, after which the number saturates and begins to decrease above 14 at.% C. The carbon atoms aggregate into clusters, mostly in the form of chains and graphene flakes, leaving less carbon to disrupt the GST matrix at higher carbon concentrations. Different degrees of carbon clustering can explain divergent experimental results for recrystallization temperature for carbon doped GST.

The response of materials to shock loading is important to planetary science, aerospace engineering, and energetic materials. Thermally activated processes, including chemical reactions and phase transitions, are significantly accelerated by energy localization into hotspots. These result from the interaction of the shockwave with the materials’ microstructure and are governed by complex, coupled processes, including the collapse of porosity, interfacial friction, and localized plastic deformation. These mechanisms are not fully understood and the lack of models limits our ability to predict shock to detonation transition from chemistry and microstructure alone. We demonstrate that deep learning can be used to predict the resulting shock-induced temperature fields in composite materials obtained from large-scale molecular dynamics simulations with the initial microstructure as the only input. The accuracy of the Microstructure-Informed Shock-induced Temperature net (MISTnet) model is higher than the current state of the art and its evaluation requires a fraction of the computation cost.

We present first-principles calculations for a system consisting of a varying number of layers of the ferroelectric semiconductor Bi(Zn, Ti)O-3 (BZT) on top of SrTiO3 (STO) substrate. To begin with, we assume that a TiO2 layer in STO is connected to a BiO layer in BZT. We show that atomic relaxation has a significant effect on the electronic properties of the heterostructure. In the absence of relaxation, due to the large polarization in BZT, a two-dimensional electron gas is formed at the interface when one unit cell of BZT is placed on top of STO. When relaxation is taken into account, the polarization in BZT is reduced in value and reversed in direction; in this case the insulator-to-metal transition occurs when the number of BZT unit cells reaches four, and a two-dimensional hole gas is formed at the interface. We also consider the case where an SrO layer in STO connects to a (Zn, Ti)O-2 layer in BZT, and show that, when three unit cells of BZT are added on top of STO, a two-dimensional electron gas is formed near the interface.

Recent research in ultrawide-bandgap (UWBG) semiconductors has focused on traditional materials such as Ga2O3, AlGaN, AlN, cubic BN, and diamond; however some materials exhibiting single perovskite structure have been known to yield bandgaps above 3.4 eV, such as BaZrO3. In this work we propose two novel materials to be added to the family of UWBG semiconductors: Ba2CaTeO6 exhibiting a double perovskite structure and Ba2K2Te2O9 with a triple perovskite structure. Using first principles hybrid functional calculations we predict the bandgaps of all the studied systems to be above 4.5 eV with strong optical absorption in the ultraviolet region. Furthermore, we show that holes have a tendency to get trapped through lattice distortions in the vicinity of oxygen atoms with average trapping energy of 0.25 eV,potentially preventing the enhancement of p-type conductivity through traditional chemical doping.