Pratik Brahma

EECS Department, University of California, Berkeley

Technical Report No. UCB/

May 1, 2025

Atomistic simulations are powerful tools for designing and understanding complex macroscopic systems with atomic-scale precision. By capturing how atomic structure influences macroscopic behavior, these methods provide predictive insights into material properties. As a result, atomistic simulations can enable breakthroughs across a wide range of scientific domains, including materials science, biophysics, and nanoelectronics. However, despite their accuracy, such simulations are computationally demanding and often fail to scale efficiently to the system sizes and timescales required for modeling practical systems.

In this work, we present two frameworks aimed at accelerating atomistic simulations: (i) hardware-accelerated Ising computing and (ii) graph neural networks (GNNs). Each framework addresses a distinct class of atomistic simulation problems. The first focuses on quantum lattice systems, which naturally map onto Ising models. We develop a synchronous Ising sampler for solving classical lattice systems and use an asynchronous Ising sampler, PASS (Parallel Asynchronous Stochastic Sampler), for solving quantum spin lattices. These hardware architectures deliver orders-of-magnitude speed-ups over GPU-based algorithms while maintaining near state-of-the-art accuracy.

The second framework tackles the more challenging problem of predicting device characteristics from atomically resolved transistor structures. We develop an end-to-end graph neural network (GNN) pipeline that learns to map microscopic atomic configurations to macroscopic transport properties. This pipeline exhibits a significant scaling advantage, achieving orders-of-magnitude acceleration in simulation time while maintaining high predictive accuracy.

Together, these approaches provide a promising path toward fast, scalable, and accurate atomistic simulations, enabling broader and more efficient application of atomistic modeling across diverse scientific and engineering domains.

Advisors: Sayeef Salahuddin

---

BibTeX citation:

@phdthesis{Brahma:31900,
    Author= {Brahma, Pratik},
    Title= {Accelerating Atomistic Simulations Using Neural Networks and Ising Computing},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Number= {UCB/},
    Abstract= {Atomistic simulations are powerful tools for designing and understanding complex macroscopic systems with atomic-scale precision. By capturing how atomic structure influences macroscopic behavior, these methods provide predictive insights into material properties. As a result, atomistic simulations can enable breakthroughs across a wide range of scientific domains, including materials science, biophysics, and nanoelectronics. However, despite their accuracy, such simulations are computationally demanding and often fail to scale efficiently to the system sizes and timescales required for modeling practical systems.

In this work, we present two frameworks aimed at accelerating atomistic simulations: (i) hardware-accelerated Ising computing and (ii) graph neural networks (GNNs). Each framework addresses a distinct class of atomistic simulation problems. The first focuses on quantum lattice systems, which naturally map onto Ising models. We develop a synchronous Ising sampler for solving classical lattice systems and use an asynchronous Ising sampler, PASS (Parallel Asynchronous Stochastic Sampler), for solving quantum spin lattices.
These hardware architectures deliver orders-of-magnitude speed-ups over GPU-based algorithms while maintaining near state-of-the-art accuracy.

The second framework tackles the more challenging problem of predicting device characteristics from atomically resolved transistor structures. We develop an end-to-end graph neural network (GNN) pipeline that learns to map microscopic atomic configurations to macroscopic transport properties. This pipeline exhibits a significant scaling advantage, achieving orders-of-magnitude acceleration in simulation time while maintaining high predictive accuracy.

Together, these approaches provide a promising path toward fast, scalable, and accurate atomistic simulations, enabling broader and more efficient application of atomistic modeling across diverse scientific and engineering domains.},
}

EndNote citation:

%0 Thesis
%A Brahma, Pratik 
%T Accelerating Atomistic Simulations Using Neural Networks and Ising Computing
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 1
%@ UCB/
%F Brahma:31900