Chapter 5 | Part 2 | Csci

High-Performance Computing (HPC)

What provides the massive computational power needed to tackle problems too large, data-intensive, or time-sensitive for standard computers and laptops?

Climate and Weather

What HPC application runs atmospheric models and analyzes massive geographic grids to predict severe weather events?

Typhoon Track and Intensity Forecasting

What is an example of an HPC application used to predict the movement and strength of typhoons?

Long-Term Climate Projections

What type of climate-related prediction uses HPC to study future climate conditions?

Genomics and Health

What HPC application handles large genomic comparisons to accelerate medical breakthroughs?

Drug Screening and Docking

What process uses HPC to examine how chemical compounds interact with target proteins?

Protein Folding Exploration

What scientific application studies how proteins fold using computational simulations?

Physics and Energy

What HPC application simulates complex physical interactions that rapidly increase in complexity?

Fluid Dynamics

What field studies the movement of air and water through computational simulations?

Fusion and Battery Materials

What are examples of research areas supported by HPC under Physics and Energy?

Molecule Design

What engineering and scientific activity uses HPC to study and develop new molecules?

Engineering and Transport

What HPC application powers time-critical simulations to ensure safety and optimize infrastructure?

Traffic Flow Optimization

What application uses HPC to improve the movement of vehicles on roads?

Storm Surge and Flood Mapping

What HPC application models the effects and geographical extent of coastal flooding and floods?

Structural Stress Simulation

What simulation evaluates how buildings and vehicles respond to extreme conditions?

AI Training

What HPC application trains next-generation artificial intelligence models using clusters of numerous GPUs?

Large Vision Models, Advanced Language Models, and Multimodal Foundation Models

What are the three examples of AI models that require HPC training?

Traditional HPC

What approach simulates complex phenomena purely through first-principles physics equations?

First-Principles Physics Equations

What equations are used by traditional HPC to simulate complex phenomena, such as fluid dynamics and partial differential equations?

AI-Assisted Scientific Discovery

What approach uses machine learning to accelerate, augment, or replace parts of the scientific workflow?

AI Surrogate Models

What machine learning models learn complex patterns from past HPC simulation data to predict outputs?

Data-Driven Predictive Power

What capability allows AI-assisted discovery to estimate simulation outputs almost instantly without recalculating full equations?

Traditional HPC

What approach requires an entire simulation to be recomputed whenever input parameters or conditions change?

AI-Assisted Discovery

What approach combines physical domain knowledge with machine learning predictions?

HPC for AI

What integration pattern uses HPC infrastructure to accelerate the training of large artificial intelligence models?

Data Parallelism

What technique allows multiple GPUs to work together on different portions of data when training a single large AI model?

HPC for AI

What integration pattern is used when training a specialized Philippine language model or satellite vision model using DOST-ASTI COARE GPU nodes?

AI for HPC

What integration pattern uses AI models to replace or mimic expensive and time-consuming HPC simulations?

Surrogate / Emulator

What is another name for an AI model trained to imitate an expensive HPC simulation?

AI for HPC

What integration pattern can provide millisecond-level predictions after learning from hundreds or thousands of HPC simulations?

Storm Surge Height Emulation

What is an example of AI for HPC that predicts storm surge heights without repeatedly running full fluid dynamics simulations?

AI in the Loop

What integration pattern combines AI and HPC into a continuous, closed-loop bidirectional workflow?

Closed-Loop Bidirectional Workflow

What type of workflow allows AI to steer HPC simulations while HPC generates new data to retrain and refine AI models?

Real-Time Monitoring and Steering

What role does AI perform in an AI in the Loop workflow while HPC simulations are running?

Dynamic Retraining and Refinement

What happens to the AI model when HPC continuously generates new simulation data in an AI in the Loop workflow?

Pre-Processing and Post-Processing

What are the two roles AI performs when it surrounds the core HPC simulation?

Pre-Processing

What stage uses machine learning to clean, filter, and merge incoming real-world observational data with simulation baselines?

Post-Processing

What stage uses AI to analyze, condense, and summarize massive amounts of raw simulation output into actionable insights?

Data Assimilation

What process merges observations and simulation data, such as satellite readings and climate model outputs?

Real-Time Satellite Readings, Ground Sensor Measurements, and Physics Model Outputs

What three data sources can be merged before running a forecast model or mapping disaster zones?

AI-Assisted Scientific Discovery

What paradigm shift changes scientific workflows from simulation-based approaches to data-driven and physics-informed approaches?

Simulation Generates Data, ML Learns

What describes the relationship between scientific simulations and machine learning in AI-assisted discovery?

Simulation and Training

What two workloads does HPC handle in AI-assisted scientific discovery?

AlphaFold2

What AI system predicted approximately 200 million proteins and helped solve a 50-year grand challenge?

200 Million Proteins

Approximately how many proteins did AlphaFold2 predict?

10 Years

How long would each protein prediction have taken through laboratory work according to the presentation?

GraphCast

What DeepMind weather emulator can produce predictions in approximately one minute instead of one hour using an ECMWF numerical model?

DeepMind

What organization developed GraphCast?

1 Minute

How long does GraphCast take to produce a weather prediction according to the presentation?

1 Hour

How long does the ECMWF numerical model take for comparison with GraphCast?

10 Days

For what forecast period does GraphCast provide more accurate predictions according to the presentation?

Surrogate Models

What AI + HPC integration pattern trains machine learning to imitate expensive simulations?

Inverse Design

What integration pattern takes a desired property as input and generates candidate materials or molecules?

Smart Sampling

What integration pattern allows AI to decide where the next simulation should be performed?

Data Assimilation

What integration pattern merges real-world observations with simulation data to improve predictions?

Anomaly Detection

What integration pattern identifies rare or unusual events in massive datasets?

Computational Fluid Dynamics (CFD)

What type of simulation is used in the example of training a surrogate model using 1,000 simulation runs?

1,000 CFD Runs

How many CFD simulations are used in the example for training a surrogate model?

1,000 Times Faster

Approximately how much faster can surrogate model predictions be compared with expensive simulations?

Battery Capacity

What desired property is used as an example in the inverse design of candidate materials?

Bayesian Optimization and Active Learning

What two methods are used in smart sampling to avoid brute-force parameter sweeps?

Satellite and Climate Model Data

What data sources are combined in the example of data assimilation?

Rogue Waves, Gravitational Waves, and Fraud in Data

What are the three examples of rare events or unusual patterns identified through anomaly detection?

Surrogate Models

What integration pattern commonly uses supervised regression algorithms to learn the relationship between input X and output y?

Supervised Regression

What machine learning approach is most commonly used for surrogate models according to the presentation?

Neural Networks / Deep Learning

What algorithm family learns input-to-output relationships and is used for surrogate models?

Gaussian Process / Kriging

What regression algorithm is commonly used for surrogate modeling and prediction?

Random Forest

What machine learning algorithm is listed as an option for surrogate models?

Generative Models

What type of algorithms are used in inverse design to generate candidates from desired properties?

GANs, VAEs, and Diffusion Models

What three algorithm families are listed for inverse design?

Generative Adversarial Networks (GANs)

What generative model is included in the inverse design algorithm family?

Variational Autoencoders (VAEs)

What generative model is included alongside GANs and diffusion models for inverse design?

Diffusion Models

What generative model is listed for producing candidates in inverse design?

Optimization

What algorithm category is associated with smart sampling?

Reinforcement Learning

What learning approach is listed alongside Bayesian Optimization and Active Learning for smart sampling?

Data Fusion

What algorithm category is associated with data assimilation?

Kalman Filter / Ensemble Kalman

What algorithm is used to merge observations and simulation data under data assimilation?

Deep Super-Resolution

What algorithm approach is listed for data assimilation?

Graph Neural Networks

What neural network architecture is listed for merging satellite and model data?

Unsupervised Learning

What machine learning category is associated with anomaly detection?

Autoencoders

What unsupervised learning models are listed for anomaly detection?

Isolation Forest

What algorithm is used to identify unusual observations under anomaly detection?

DBSCAN and K-Means

What two clustering algorithms are listed for anomaly detection?

Convolutional Neural Networks (CNN)

What neural network architecture is listed for anomaly detection?

Brute-Force Sweep

What traditional approach divides a parameter space into a rigid, uniform grid and runs a simulation at every point?

Smart Sampling

What AI-driven approach dynamically selects where to simulate next based on previous results?

Parameter Space

What is divided into a uniform grid during a brute-force sweep?

100,000 Simulations

How many simulations are required when five variable parameters each have ten possible values in a brute-force sweep?

Bayesian Optimization

What method helps evaluate a parameter space and identify promising points for simulation?

Active Learning

What method allows an AI system to select the most informative points for the next simulation?

Probabilistic Surrogate Model

What model is built by smart sampling to selectively target points with high information gain or potential performance peaks?

Information Gain

What does smart sampling seek to maximize when selecting the next simulation points?

Performance Peak

What potential result does smart sampling target when choosing where to run the next simulation?

Speed, Scale, and Insight

What are the three main benefits of AI-assisted discovery mentioned in the evaluation section?

Hallucination

What risk occurs when AI generates physics that appears correct but violates actual physical laws?

Bias

What risk occurs when an AI model's predictions are affected by biases in its training data?

Black Box

What risk describes an AI system whose decision-making process cannot be explained?

Garbage In, Garbage Out

What principle describes the limitation of AI systems that depend on the quality of their input data?

Extrapolation Failure

What limitation occurs when an AI model makes predictions outside the data on which it was trained?

Physics

What can AI not replace or independently discover according to the presentation?