AI/ML Accelerating HPC Scientific Applications is the integration of neural networks and machine learning methods into high-performance computing workflows — replacing or augmenting expensive physics-based simulations with learned surrogate models, neural operators, and AI-driven force fields that can be 100-10000× faster while maintaining sufficient accuracy for scientific discovery, fundamentally changing the computational economics of climate modeling, drug discovery, materials science, and nuclear stockpile simulation.

Neural Surrogate Models

Replace expensive simulation runs with fast ML approximations:

• Training: run the expensive simulator for hundreds/thousands of input configurations, train ML model to approximate input→output mapping.
• Inference: new inputs evaluated in milliseconds instead of hours.
• Applications: aerodynamic drag prediction (CFD surrogate), nuclear cross-section interpolation, turbine design optimization.
• Uncertainty quantification: surrogate must indicate when it is out-of-distribution (Gaussian process surrogate provides variance estimate; deep ensembles for neural surrogates).

Physics-Informed Machine Learning

• PINNs (Physics-Informed Neural Networks): loss function includes PDE residual (forces solution to satisfy governing equations), handles inverse problems (infer parameters from measurements).
• Fourier Neural Operator (FNO): learns operator (function space → function space), applied to Navier-Stokes, weather, seismic. 1000× faster than FEM for Navier-Stokes at same resolution.
• DeepONet: universal approximation theorem for operators, two-branch architecture.
• Neural ODE: continuous-depth model (ODE system learned by neural net), used for time series and latent dynamics.

ML Turbulence Modeling

Reynolds-averaged Navier-Stokes (RANS) requires closure model for turbulence (k-ε, k-ω models are empirical). ML turbulence:

• Train neural network to predict Reynolds stress tensor from flow features.
• Improves accuracy over empirical closures for complex geometries.
• Embedded in CFD solver (ANSYS Fluent, OpenFOAM) via neural network inference.

ML Force Fields for Molecular Dynamics

Ab initio MD (AIMD) computes quantum mechanical forces per step: O(N³) — limited to 100s of atoms for picoseconds.

• NNP (Neural Network Potentials): train on DFT force/energy labels, infer forces in O(N). ANI-2x, NequIP, MACE, SevenNet.
• Accuracy: within 1 kcal/mol of DFT for in-distribution configurations.
• Speed: 1000× faster than AIMD, enables million-atom systems, microsecond timescales.
• Applications: protein folding kinetics, battery electrolyte stability, catalyst activity prediction.

AI-Driven Adaptive Mesh Refinement (AMR)

• RL agent decides where to refine mesh based on local error estimate.
• Learns to allocate resolution budget optimally for given physics.
• Applied to plasma physics (fusion) simulations.

Generative AI for Scientific Data

• Data augmentation: generate synthetic training data for rare events (extreme weather, rare chemical configurations).
• Scientific image synthesis: generate synthetic microscopy images (electron microscopy) for segmentation model training.
• Inverse design: generate molecular structures with target properties (drug-likeness, band gap).

AI/ML for HPC is the transformative fusion of data-driven learning with physics-based simulation that amplifies the scientific output of supercomputing investments — enabling researchers to explore vast parameter spaces, discover new materials, and model complex phenomena at scales and speeds that pure simulation or pure ML alone cannot achieve.