Jack T. Beerman

Physical AI · Physics-aware deep learning

Jack T. Beerman

I build neural networks that learn how the physical world moves.

My models forecast shock waves, floods and deforming solids. They carry derivative operators and numerical time integrators inside the architecture, and they adapt their resolution to wherever the physics gets sharp.

See the research ↓ GitHub ↗ Google Scholar ↗ Hugging Face ↗ LinkedIn ↗
Illustration: blast waves on an unstructured mesh, colored by pressure. Cells refine where the gradient is steep. t = 0.42

Research

Physics-aware deep learning

Numerical solvers are accurate and slow. Generic neural networks are fast, but they drift once a rollout runs long or leaves the training data. Physics-aware recurrent convolutional networks (PARC) split the work: a network learns the dynamics that are hard to write down, and explicit derivatives and time integrators handle the parts we already know.

PhD dissertation · 2026

Adaptive Discretization in Physics-Aware Deep Learning

Most physics-aware networks borrow off-the-shelf architectures with fixed grids and constant time steps, which break down in strongly nonlinear regimes. My dissertation builds adaptive discretization into the network itself, so it can change its spatial and temporal resolution as the solution evolves. The work moves from rigid Eulerian grids to hybrid Lagrangian–Eulerian refinement and then to fully unstructured domains.

  • SpaceDeformable convolutionsKernels that follow advecting features such as shocks and reaction fronts.
  • TimeVariable time integrationOne trained model steps forward at irregular Δt.
  • GeometryGraph networksPhysics-aware learning on unstructured and moving meshes.
Read the dissertation ↗

Scientific Reports · 2026 · First author

G-PARC: physics-aware graph networks on unstructured meshes

Convolutional PARC only runs on uniform Cartesian grids. G-PARC moves it onto graphs, so it works on the irregular meshes that engineering simulations actually use, including meshes that move as a structure deforms. Spatial derivatives come from moving least squares (MLS) kernels on the mesh stencil, and a numerical integrator advances the state.

604–2,382×
higher throughput than neural-operator baselines (GINO, GNO)
Any Δt
one trained model handles arbitrary timesteps at inference
3 domains
elastoplastic impact, planar shock waves, river flood forecasting
Paper ↗ arXiv ↗ Code ↗ Weights & data ↗

Architecture · differentiate, then integrate

  1. x(t) node and edge state on the mesh
  2. GATConv
    learned features
    MLS operators
    ∇ · ∇² · strain
  3. SPADE + FiLM fuse physics, condition on Δt
  4. ∂x/∂t → RK4 · Heun · Euler
  5. x(t + Δt)
    ↺ repeat for each step of the rollout

arXiv preprint · 2026 · First author

Size is not the solution: deformable convolutions for physics-aware deep learning

In most of AI, harder problems get bigger models. For physics, scaling gives diminishing returns. D-PARC replaces the fixed 3×3 kernels in PARCv2's U-Net with deformable ones that learn a sampling offset at every pixel. Across Burgers' equation, Navier–Stokes and reactive flows, it beats a substantially larger PARCv2.

The learned kernels anti-cluster into an "active filtration" pattern. They concentrate on high-strain regions and coarsen elsewhere, much like adaptive mesh refinement in computational mechanics. The model learned this without being told to.

arXiv ↗ Code ↗

Sampling pattern · one 3×3 kernel

standard: fixed square deformable: learned offsets follow the front

Each sample point moves by a learned (Δx, Δy) and is read by bilinear interpolation, so the receptive field bends toward shocks, vortices and reaction fronts.

International Journal of Multiphase Flow · 2024 · Co-author

PARC for multiphase compressible flows

Physics-aware recurrent convolutions built around the advection–diffusion–reaction equation, applied to shocks interacting with particles and to reaction fronts. The benchmarks are Burgers' equation, flow past a cylinder and shock–particle interaction across Mach numbers.

Paper ↗

Projects

Always building something new

I learn by building. Each of these started as a question outside my dissertation: how to forecast motion when data arrives irregularly, where a small fast model fits inside a larger AI system, how different RL algorithms learn to race. Each one taught me a new corner of the field.

2026 · GNN · Transformer · Probabilistic

Vessel trajectory forecasting from irregular AIS

Predicts where a ship will be as a distribution of plausible paths, using raw position reports at the irregular times they actually arrive. A heterogeneous graph network encodes nearby traffic and coastline, a time-aware causal transformer reads the history, and a sampling head trained with an energy-score loss draws many futures per pass. The design follows GraphCast and WeatherNext, and the repo includes an open benchmark.

Code ↗

2026 · LLM systems · Evaluation

Fast classifiers inside an LLM pipeline

Tests where a fast, cheap classification model earns its place in a real decision system. The classifier makes typed judgments in about 300 ms and triages roughly 10,000 prediction markets before any expensive LLM research runs. The measurements and findings are written up in the repo.

Code ↗

2024 · Reinforcement learning · PPO · SAC

Autonomous racing with AWS DeepRacer

Compared PPO and SAC for training a racing agent, then iterated on reward shaping. The best agent finished 53rd of 194 in the U.S. in the April 2024 race. Team project with Thomas Hammons, Conor McLaughlin and Luke Rohlwing.

Code ↗

Publications

Papers

Full list with citation counts on Google Scholar.

Physical AI and scientific machine learning

  • 2026

    Adaptive Discretization in Physics-Aware Deep Learning PhD dissertation

    J. T. Beerman. Advisor: S. Baek.

    School of Data Science, University of Virginia

    Dissertation

  • 2026

    G-PARC: Graph-Physics Aware Recurrent Convolutional neural networks for spatiotemporal dynamics on unstructured meshes

    J. T. Beerman, T. J. Abele, M. Taghizadeh, A. Davis, Z. J. Gray, N. Alemazkoor, et al.

    Scientific Reports

    PaperarXivCodeWeights

  • 2026

    Size is Not the Solution: Deformable Convolutions for Effective Physics Aware Deep Learning

    J. T. Beerman, S. Roy, H. S. Udaykumar, S. S. Baek

    arXiv:2601.11657

    arXivCode

  • 2024

    Physics-aware recurrent convolutional neural networks for modeling multiphase compressible flows

    X. Cheng, P. C. H. Nguyen, P. K. Seshadri, M. Verma, Z. J. Gray, J. T. Beerman, H. S. Udaykumar, S. S. Baek

    International Journal of Multiphase Flow, vol. 177

    Paper

Simulation, agent-based modeling and security

  • 2024

    Accelerating Hybrid Agent-Based Models and Fuzzy Cognitive Maps: How to Combine Agents who Think Alike?

    P. J. Giabbanelli, J. T. Beerman

    Winter Simulation Conference (WSC)

    arXiv

  • 2023

    A Review of Colonial Pipeline Ransomware Attack Most cited

    J. Beerman, D. Berent, Z. Falter, S. Bhunia

    IEEE/ACM CCGrid Workshops (CCGridW)

    Paper

  • 2023

    A framework for the comparison of errors in agent-based models using machine learning

    J. T. Beerman, G. G. Beaumont, P. J. Giabbanelli

    Journal of Computational Science, vol. 72

    Paper

  • 2023

    On the Necessity of Human Decision-Making Errors to Explain Vaccination Rates for Covid-19: An Agent-Based Modeling Study

    J. T. Beerman, G. G. Beaumont, P. J. Giabbanelli

    Annual Modeling and Simulation Conference (ANNSIM)

    Paper

  • 2023

    Accelerating Agent-Based Models and Fuzzy Cognitive Maps via CUDA

    K. Ghumrawi, K. Ha, J. Beerman, J. D. Rudie, P. J. Giabbanelli

    Hawaii International Conference on System Sciences (HICSS)

    Paper

  • 2023

    To Err Is Human: The Effect of Mistakes in Social Simulations MS thesis

    J. T. Beerman

    Miami University

  • 2022

    A Scoping Review of Three Dimensions for Long-Term COVID-19 Vaccination Models: Hybrid Immunity, Individual Drivers of Vaccinal Choice, and Human Errors

    J. T. Beerman, G. G. Beaumont, P. J. Giabbanelli

    Vaccines, 10(10), 1716

    Paper

About

Machine learning for physical systems

I'm a machine learning researcher working on physical AI: neural networks that model how physical systems evolve in space and time. I focus on the architecture itself. That means deciding where the network should look, how it should step through time, and what geometry it should run on, so it stays accurate where the physics is hardest.

I completed my PhD in Data Science at the University of Virginia, advised by Stephen Baek, where I developed G-PARC and D-PARC. Before that I earned a BS and MS in Computer Science at Miami University. There I worked with Philippe Giabbanelli on agent-based simulation, GPU acceleration and machine learning for modeling human behavior.

I'm interested in collaborating on neural surrogates, graph networks for simulation and physical AI more broadly. The fastest way to reach me is LinkedIn or GitHub.

Portrait of Jack T. Beerman
  • PhD · Data ScienceUniversity of VirginiaDissertation: Adaptive Discretization in Physics-Aware Deep Learning (2026)
  • MS · Computer ScienceMiami UniversityThesis: To Err Is Human: The Effect of Mistakes in Social Simulations
  • BS · Computer ScienceMiami University
  • Research areasPhysics-aware deep learning · Graph neural networks · Scientific ML · Network science