Beyond DFT’s Computational Limits: EquFlashV2 and the Future of AI-Driven Semiconductor Materials Simulation

From atoms to device performance

What if we could verify the performance and durability of new semiconductor materials before fabricating a single sample? Atomic-level simulation allows us to explore material changes and defects that are difficult to observe directly.

As semiconductor technology advances, minute differences in material composition, atomic defects, and interface structures can significantly affect key device characteristics, such as leakage current, reliability, and thermal transport. However, observing and optimizing these atomic-scale phenomena through experiments alone remains challenging.

Atomic-level simulations allow researchers to evaluate the properties of candidate materials before actual fabrication and analyze how defects and interfaces affect device performance. Today, these simulations are widely used across the entire semiconductor R&D—from materials exploration and interface engineering to process optimization and reliability analysis—helping reduce trial-and-error and providing atomic-scale insights that are difficult to obtain through experiments alone.

The bottleneck: accuracy vs. computational cost

While the scope of atomic-level simulation has expanded rapidly, higher accuracy typically requires more time and computing resources. This limitation becomes even more pronounced when dealing with larger and more complex atomic systems.

First-principles methods such as density functional theory(DFT) are widely used to analyze atomic structures, stability, and material behavior with high accuracy. However, as systems grow larger and calculations become more complex, computational cost increases rapidly, making it impractical to study large numbers of candidate materials, complex atomic configurations, or large-scale simulation scenarios.

Machine Learning Force Fields(MLFFs): breaking the computational barrier

Machine Learning Force Fields(MLFFs) offer an effective way to address these computational limitations. In many first-principles methods, the cost of electronic structure calculations scales approximately as O(N³) with system size, making repeated calculations for large supercells, complex interfaces, defects, and amorphous structures increasingly difficult.

MLFFs are trained to approximate energy, force, and stress at the first-principles level based on atomic structures, reducing computational scaling to approximately O(N). This enables the exploration of larger atomic systems and a wider array of structures and conditions.

However, speed is not the only metric for MLFF performance. In practical atomistic simulations, models are used iteratively during processes like structure relaxation and property prediction. Therefore, stable performance throughout these iterative processes is just as critical as single-point prediction accuracy. Matbench Discovery reflects these real-world requirements by comprehensively evaluating stability prediction, structure relaxation, and phonon-based property prediction, providing a standard for comparing the practical performance of MLFFs.

EquFlashV2: integrating physical symmetry with computational efficiency

EquFlashV2 is an E(3)-equivariant MLFF developed to accelerate atomic-level material simulations.
E(3)-equivariance is a design principle ensuring that physical predictions remain consistent even if the atomic structure is rotated, translated, or reflected. For example, while the energy remains the same regardless of the structure's orientation, directional physical quantities—such as force—must change in tandem with the structural transformation.

EquFlashV2 represents atomic structures as graphs and learns inter-atomic interactions through equivariant message passing. Based on this, it predicts atomic-level physical quantities, including energy, force, and stress, required for structure relaxation and subsequent property evaluation. By combining Clebsch–Gordan tensor product-based message passing with quadratic gate activation functions, the model captures interactions that depend on atomic distances and directions while preserving the required physical symmetries.

A key strength of EquFlashV2 lies in its model architecture, which is optimized for efficient equivariant operations on GPUs. By applying a Transformer-style block architecture that use tensor-product-based operations instead of attention mechanisms, the model can expand its parameters and layer count while limiting additional training and inference overhead, thereby enabling greater representational capacity than previous models.

This design makes large-scale atomistic simulations more practical, reducing the computational burden of evaluating structural stability, defects, and interfaces—tasks that were previously impractical using only first-principles calculations. Consequently, it enables the rapid exploration of more candidate materials and the efficient comparison of various material structures and semiconductor process conditions.

From validation to real-world semiconductor R&D

Based on the Combined Performance Score(CPS), EquFlashV2 achieved state-of-the-art(SOTA) performance on the Matbench Discovery benchmark¹. This result demonstrates its competitiveness as a machine learning force field for atomic-level simulation and underscores the potential of AI-driven materials research.

Currently, EquFlashV2 is being utilized in our device development and material screening for process development. Moving forward, we plan to expand its application to high-throughput virtual screening to explore, optimize, and select promising candidate materials for next-generation memory and logic devices.

Samsung Electronics will continue to advance EquFlashV2 and related technologies and integrate them with open-source platforms, with the goal of establishing MLFFs as practical R&D tools for accelerating next-generation semiconductor materials development.

References