CHAPTER 5.2 - High-Performance Computing & AI Assisted Scientific Discovery

Too large, data intensive, time sensitive

HPC changes science by providing the massive computational power required to tackle problems that are ___ _____, ____ _________, or ____ _________ for standard computers and laptops.

Climate and weather, genomics and health, physics and energy, engineering and transport, ai training

5 fields where HPC changes science (CGPEA)

Climate and Weather

field where hpc changes science.

What it does: Runs atmospheric models and analyzes massive geographic grids to predict severe weather events.

Example: Typhoon track & intensity forecasting or long-term climate projections. Because weather systems involve vast geographical areas and rapid changes, forecasters need fast, massive grid calculations to issue timely safety warnings.

Genomics and Health

field where hpc changes science.

What it does: Handles big data genomic comparisons to accelerate medical breakthroughs.

Example: Drug screening & docking or protein folding exploration. Simulating how thousands of chemical compounds interact with target proteins takes massive computational parallelism, turning years of lab work into days.

Physics and Energy

field where hpc changes science.

What it does: Simulates complex physical interactions that scale rapidly in complexity.

Example: Fluid dynamics for air and water, studying fusion and battery materials, or the design of new molecules.

Engineering and Transport

field where hpc changes science.

What it does: Powers time-critical simulations to ensure safety and optimize infrastructure.

Example: Traffic flow optimization, storm surge and flood mapping, or structural stress simulation for buildings and vehicles under extreme conditions

AI Training

field where hpc changes science.

What it does: Trains next generation artificial intelligence models that require clusters of numerous GPUs working in tandem.

Example: Large vision models, advanced language models, and multi-modal foundation models.

Traditional HPC

Simulates complex phenomena purely from first-principles physics equations (e.g., fluid dynamics, partial differential equations).

Limitation: Highly accurate, but extremely slow and computationally expensive. Every change in input parameters or conditions requires computing the entire simulation from scratch.

AI-Assisted Discovery

Uses machine learning models (also called AI surrogate models) to learn complex patterns from past HPC simulation data.

Advantage: Combines physical domain knowledge with data-driven predictive power to estimate outputs almost instantly without recalculating full equations each time.

AI surrogate models

machine learning models are also called what?

Data prep, data interpretation

The core concept of the AI+HPC patterns is AI surrounds the core HPC simulation by handling ____ ____ before computation and ____ ______________ after computation. *HPC for truth and detail, AI for speed and pattern*.

HPC for AI, AI for HPC, AI in the Loop, AI around HPC, Digital Twin

5 patterns of AI+HPC Integration (HAAAD)

HPC for AI (Surrogate Model)

pattern of AI+HPC. Using high-performance computing (HPC) to train large AI models faster by distributing tasks across multiple GPUs.

HPC for AI (Surrogate Model)

pattern of AI+HPC. Best used when training large AI models with massive datasets and limited time.

HPC for AI (Surrogate Model)

pattern of AI+HPC. Main goal is to reduce AI training time from weeks to hours.

AI for HPC (Inverse Design)

pattern of AI+HPC. Using AI models to imitate or replace expensive and time-consuming HPC simulations.

AI for HPC (Inverse Design)

pattern of AI+HPC. Best used when fast predictions and multiple what-if scenarios are needed.

AI for HPC (Inverse Design)

pattern of AI+HPC. Main goal is to reduce simulation time from hours to milliseconds.

AI in the Loop (Smart Sampling)

pattern of AI+HPC. combines both into a continuous, closed-loop bidirectional workflow. Instead of AI acting as a static substitute, AI and HPC run concurrently: AI monitors and steers the HPC simulation in real time, while HPC continuously generates new data to retrain and refine the AI model dynamically

Brute-Force Sweep

under pattern 3, AI in the loop. Scientists divide a parameter space into a rigid, uniform grid and run an HPC simulation for every single point, regardless of whether that specific area is interesting or physically relevant. It wastes massive amounts of computational power and time. If a simulation has 5 variable parameters and each has 10 possible values, a brute-force grid requires running 105(100,000) simulations. For heavy physics models taking hours each, this is entirely unfeasible.

Smart Sampling

under pattern 3, AI in the loop. Instead of sampling blindly across a grid, AI dynamically decides where to simulate next based on previous results. By combining Bayesian optimization and active learning, the system evaluates the parameter space, builds a probabilistic surrogate model, and selectively targets only the points that offer the highest information gain or potential performance peak.

AI Around HPC (Data Assimilation)

pattern of AI+HPC. ML merges observations + simulation, or analyzes huge HPC outputs - cleans before and summarizes after.

Digital Twin (Anomaly Detection)

pattern of AI+HPC. Live physical system + HPC simulation + Al continuousty updating each other in closed loop - real-time. Find rare events in PB outputs - rogue wave, gravitational wave, fraud in data

Neural Networks/Deep Learning, Gaussian Process/Kriging, Random Forest

3 common ML algorithms under pattern 1 (NGR). Surrogate → Supervised Regression

GANs, VAEs, Diffusion

3 common ML algorithms under pattern 2 (GVD). Inverse → Generative. Desired property → generates candidate.

Bayesian Optimization, Active Learning, Reinforcement Learning

3 common ML algorithms under pattern 3 (BAR). Smart Sampling → Optimization. Picks next best simulation.

Kalman Filter/Ensemble Kalman, Deep Super-Resolution, Graph Neural Networks

3 common ML algorithms under pattern 4 (KDG). Data Assimilation → Data Fusion. Merges satellite and model data.

Autoencoders, Isolation Forest, DBSCAN/k-means, CNN Detection

4 common ML algorithms under pattern 5 (AIDC). Anomaly → Unsupervised. Learns normal patterns and flags rare events.

Speed, Scale, Insight

3 main benefits of AI assisted approaches (SSI)

Hallucination, Bias, Black Box

3 common risks of AI assisted approaches. (HBB)

Data Dependency, Extrapolation Failure, Physics Dependency

3 major limitations of AI assisted approaches. (DEP)