arXiv subjects

Anup Pandey

Publications and source records attributed to Anup Pandey.

8 recordsLinked to original sources

arXiv2026

PowerModelsGAT-AI: Physics-Informed Graph Attention for Multi-System Power Flow with Continual Learning

Solving the alternating current power flow equations in real time is essential for secure grid operation, yet classical Newton-Raphson solvers can be slow under stressed conditions. Existing graph neural networks for power flow are typically trained on a single system and often degrade on different systems. We present PowerModelsGAT-AI, a physics-informed graph attention network that predicts bus voltages and generator injections. The model uses bus-type-aware masking to handle different bus types and balances multiple loss terms, including a power-mismatch penalty, using learned weights. We evaluate the model on 14 benchmark systems (4 to 6,470 buses) and train a unified model on 13 of these under N-2 (two-branch outage) conditions, achieving an average normalized mean absolute error of 0.89% for voltage magnitudes and R^2 > 0.99 for voltage angles. We also show continual learning: when adapting a base model to a new 1,354-bus system, standard fine-tuning causes severe forgetting with error increases exceeding 1000% on base systems, while our experience replay and elastic weight consolidation strategy keeps error increases below 2% and in some cases improves base-system performance. Interpretability analysis shows that learned attention weights correlate with physical branch parameters (susceptance: r = 0.38; thermal limits: r = 0.22), and feature importance analysis supports that the model captures established power flow relationships.

eess.SY↗

arXiv2024

Best of Both Worlds: Enforcing Detailed Balance in Machine Learning Models of Transition Rates

The slow microstructural evolution of materials often plays a key role in determining material properties. When the unit steps of the evolution process are slow, direct simulation approaches such as molecular dynamics become prohibitive and Kinetic Monte-Carlo (kMC) algorithms, where the state-to-state evolution of the system is represented in terms of a continuous-time Markov chain, are instead frequently relied upon to efficiently predict long-time evolution. The accuracy of kMC simulations however relies on the complete and accurate knowledge of reaction pathways and corresponding kinetics. This requirement becomes extremely stringent in complex systems such as concentrated alloys where the astronomical number of local atomic configurations makes the a priori tabulation of all possible transitions impractical. Machine learning models of transition kinetics have been used to mitigate this problem by enabling the efficient on-the-fly prediction of kinetic parameters. In this study, we show how physics-informed ML architectures can exactly enforce the detailed balance condition, by construction. Using the diffusion of a vacancy in a concentrated alloy as an example, we show that such ML architectures also exhibit superior performance in terms of prediction accuracy, demonstrating that the imposition of physical constraints can facilitate the accurate learning of barriers at no increase in computational cost.

cond-mat.mtrl-sci↗

arXiv2022

Machine learning interatomic potential for high throughput screening and optimization of high-entropy alloys

We have developed a machine learning-based interatomic potential (MLIP) for the quaternary MoNbTaW (R4) and quinary MoNbTaTiW (R5) high entropy alloys (HEAs). MLIPs enabled accurate high throughput calculations of elastic and mechanical properties of various non-equimolar R4 and R5 alloys, which are otherwise very time consuming calculations when performed using density functional theory (DFT). We demonstrate that the MLIP predicted properties compare well with the DFT results on various test cases and are consistent with the available experimental data. The MLIPs are also utilized for high throughput optimization of non equimolar R4 candidates by guided iterative tuning of R4 compositions to discover candidate materials with promising hardness-ductility combinations. We also used this approach to study the effect of Ti concentration on the elastic and mechanical properties of R4, by statistically averaging the properties of over 100 random structures. MLIP predicted hardness and bulk modulus of equimolar R4 and R5 HEAs are validated using experimentally measured Vickers hardness and modulus. This approach opens a new avenue for employing MLIPs for HEA candidate optimization.

cond-mat.mtrl-sci↗

arXiv2020

Machine learning enabled surrogate crystal plasticity model for spatially resolved 3D orientation evolution under uniaxial tension

We present a novel machine learning based surrogate modeling method for predicting spatially resolved 3D microstructure evolution of polycrystalline materials under uniaxial tensile loading. Our approach is orders of magnitude faster than the existing crystal plasticity methods enabling the simulation of large volumes that would be otherwise computationally prohibitive. This work is a major step beyond existing ML-based modeling results, which have been limited to either 2D structures or only providing average, rather than local, predictions. We demonstrate the speed and accuracy of our surrogate model approach on experimentally measured microstructure from high-energy X-ray diffraction microscopy of a face-centered cubic copper sample, undergoing tensile deformation.

cond-mat.mtrl-sci↗

arXiv2017

Evolution of amorphous carbon across densities: an inferential study

In this paper, we offer large and realistic models of amorphous carbon spanning densities from 0.95 g/cm3 to 3.5 g/cm3 . The models are designed to agree as closely as possible with experimental diffraction data while simultaneously attaining a local minimum of a density functional Hamilto- nian. The structure varies dramatically from interconnected wrapped and defective sp2 sheets at 0.95 g/cm3 to a nearly perfect tetrahedral topology at 3.5 g/cm3 . Force Enhanced Atomic Refine- ment (FEAR) was used and is shown here to be computationally superior and more experimentally realistic than conventional ab initio melt quench methods. We thoroughly characterize our models by computing structural, electronic and vibrational spectra. The vibrational density of states of the 0.95 g/cm3 model is strikingly similar to monolayer amorphous graphene. Our sp2 /sp3 ratios are close to experimental predictions where available, a consequence of compelling a satisfactory fit for pair correlation function.

cond-mat.dis-nn↗

arXiv2016

Ab initio model of amorphous zinc oxide (a-ZnO) and a-X_0.375 Z_0.625 O (X=Al, Ga and In)

Density functional theory (DFT) calculations are carried out to study the structure and electronic structure of amorphous zinc oxide (a-ZnO). The models were prepared by the "melt-quench" method. The models are chemically ordered with some coordination defects. The effect of trivalent dopants in the structure and electronic properties of a-ZnO is investigated. Models of a-X_0.375 Z_0.625 O (X=Al, Ga and In) were also prepared by the "melt- quench" method. The trivalent dopants reduce the four-fold Zn and O, thereby introducing some coordination defects in the network. The dopants prefer to bond with O atom. The network topology is discussed in detail. Dopants reduce the gap in EDOS by producing defect states minimum while maintaining the extended nature of the conduction band edge.

cond-mat.mtrl-sci↗

arXiv2016

Realistic inversion of diffraction data for an amorphous solid: the case of amorphous silicon

We apply a new method "force enhanced atomic refinement" (FEAR) to create a computer model of amorphous silicon (a-Si), based upon the highly precise X-ray diffraction experiments of Laaziri et al. The logic underlying our calculation is to estimate the structure of a real sample a-Si using experimental data and chemical information included in a non-biased way, starting from random coordinates. The model is in close agreement with experiment and also sits at a suitable minimum energy according to density functional calculations. In agreement with experiments, we find a small concentration of coordination defects that we discuss, including their electronic consequences. The gap states in the FEAR model are delocalized compared to a continuous random network model. The method is more efficient and accurate, in the sense of fitting the diffraction data than conventional melt quench methods. We compute the vibrational density of states and the specific heat, and find that both compare favorably to experiments.

cond-mat.mtrl-sci↗

arXiv2016

Inversion of diffraction data for amorphous materials

The general and practical inversion of diffraction data-producing a computer model correctly representing the material explored - is an important unsolved problem for disordered materials. Such modeling should proceed by using our full knowledge base, both from experiment and theory. In this paper, we describe a robust method to jointly exploit the power of ab initio atomistic simulation along with the information carried by diffraction data. The method is applied to two very different systems: amorphous silicon and two compositions of a solid electrolyte memory material silver-doped GeSe3 . The technique is easy to implement, is faster and yields results much improved over conventional simulation methods for the materials explored. By direct calculation, we show that the method works for both poor and excellent glass forming materials. It offers a means to add a priori information in first principles modeling of materials, and represents a significant step toward the computational design of non-crystalline materials using accurate interatomic interactions and experimental information.

cond-mat.mtrl-sci↗