Why Science? | Institute of High Performance Computing

A*STAR says IHPC was established in August 1998 to lead computational modelling, simulation and AI. Its scope spans physical, engineered and human systems. The common thread is not one supercomputer or algorithm; it is the disciplined use of computation to answer domain questions with testable evidence.

Computational science turns equations, algorithms and data into testable representations of physical or human systems; validation decides where those representations can be trusted. The chapter develops institutional literacy. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Map three IHPC domains to the models, data and reviewers each needs. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: An institute capability does not establish the result of an unnamed project.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Next chapter

Models keep selected features of a system and leave others out. A useful model begins with a question, scale and intended decision, then states assumptions and variables accordingly. Complexity is valuable only when it improves the answer relative to a simpler baseline.

High-performance computing gains value through parallel work, but speed-up claims require the same task, hardware, precision, energy and accuracy boundaries. The chapter develops model formulation. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Write a model card with purpose, included processes and omitted processes. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Every model has a domain in which its assumptions can fail.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Differential equations, probability models or agent rules express how quantities change and interact. The equation choice carries scientific assumptions about conservation, randomness and causality. Parameters may come from measurement, literature or calibration and should keep their sources and units.

Physics-based and data-driven models can complement one another when constraints, training evidence, uncertainty and failure cases remain inspectable. The chapter develops mathematical modelling. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Annotate one equation with variable meanings, units and evidence sources. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A mathematically valid equation may still be a poor representation of reality.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Simulations divide space, time or state into manageable pieces. Finer resolution can reduce some errors while increasing cost and sometimes instability. Grid, time step and solver choices need convergence tests so visual detail is not mistaken for physical accuracy.

Human-centred modelling needs consent, representative evidence and safeguards because a useful population pattern may still fail or harm an individual. The chapter develops numerical analysis. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Run a three-level resolution check and compare a meaningful output. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A high-resolution image does not prove a converged or validated simulation.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

HPC uses parallel and heterogeneous processors to tackle large computations. Algorithms expose tasks that can run together, while communication, memory movement and serial work limit speed-up. Performance reports should specify hardware, precision, problem size and accuracy.

Computational science turns equations, algorithms and data into testable representations of physical or human systems; validation decides where those representations can be trusted. The chapter develops parallel-computing literacy. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a strong-scaling table with runtime, processors and efficiency. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Twice the processors rarely produces exactly twice the speed.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC names multicore CPUs, GPUs, FPGAs and ASICs among its architectures. Each balances programmability, parallel throughput, memory and energy differently. Porting code can change numerical precision or order of operations, so equivalent scientific output must be checked rather than assumed.

High-performance computing gains value through parallel work, but speed-up claims require the same task, hardware, precision, energy and accuracy boundaries. The chapter develops computer-architecture reasoning. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Compare one workload across two architectures with accuracy and energy metrics. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A faster benchmark on one task does not make hardware universally superior.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Results depend on source code, compiler, libraries, container or environment, random seeds, hardware and input data. Version control and automated workflows make the computation inspectable. Exact bitwise repetition may be unrealistic across platforms, so teams should define scientifically meaningful tolerances.

Physics-based and data-driven models can complement one another when constraints, training evidence, uncertainty and failure cases remain inspectable. The chapter develops computational reproducibility. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build a reproducibility manifest and rerun a benchmark from a clean environment. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Sharing code alone may omit data, dependencies or execution context.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Verification asks whether equations were solved correctly; validation asks whether the model represents the intended real system well enough. Benchmarks, independent observations and out-of-sample tests reveal different errors. Calibration should not reuse all evidence needed for validation.

Human-centred modelling needs consent, representative evidence and safeguards because a useful population pattern may still fail or harm an individual. The chapter develops verification and validation. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Separate code verification, parameter calibration and external validation. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Matching calibration data is not independent validation.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Input measurements, parameters, model structure and numerical approximation all contribute uncertainty. Ensembles and sensitivity analysis show how these sources affect outputs. Reporting only a central prediction can hide decisions that are fragile to plausible changes.

Computational science turns equations, algorithms and data into testable representations of physical or human systems; validation decides where those representations can be trusted. The chapter develops uncertainty quantification. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create an uncertainty budget and vary the three most influential assumptions. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A narrow numerical interval may omit structural or scenario uncertainty.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC describes AI that uses physical knowledge, simulations and data. Conservation laws or governing equations can guide learning and improve plausibility when data are scarce. The constraints themselves may be incomplete, and performance still needs tests on new conditions and failure modes.

High-performance computing gains value through parallel work, but speed-up claims require the same task, hardware, precision, energy and accuracy boundaries. The chapter develops physics-based machine learning. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Compare a data-only baseline with a physics-informed model on extrapolation. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Physical constraints do not guarantee correct predictions or complete physics.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC studies data-efficient learning and model compression for constrained devices. Efficiency should include accuracy, latency, memory and energy on the target hardware. A smaller model can widen access, but compression may change calibration or subgroup performance.

Physics-based and data-driven models can complement one another when constraints, training evidence, uncertainty and failure cases remain inspectable. The chapter develops efficient-AI evaluation. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Benchmark the original and compressed models on task, device and subgroups. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Parameter count alone is not a complete efficiency measure.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Text, images, sensor streams and domain knowledge can complement one another. Alignment errors, missing modalities and shortcut learning may make fusion worse than strong single-mode baselines. Evaluation should test each modality alone, together and under realistic missing-data conditions.

Human-centred modelling needs consent, representative evidence and safeguards because a useful population pattern may still fail or harm an individual. The chapter develops multimodal evaluation. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build an ablation table and a missing-modality stress test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: More modalities do not automatically create more reliable understanding.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Transport, cities, supply chains and populations contain many interacting agents and feedbacks. Simulations can compare scenarios when rules and data are explicit. They do not forecast a single inevitable future; behavioural adaptation and policy response can change the system being modelled.

Computational science turns equations, algorithms and data into testable representations of physical or human systems; validation decides where those representations can be trusted. The chapter develops complex-systems thinking. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Run scenario ranges and identify feedbacks that reverse an initial result. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A scenario is a conditional exploration, not a prediction guarantee.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC describes distributed systems, federated learning, secure multiparty computation and trusted collaboration. Keeping data local may reduce transfer risk, while updates, metadata and identities still require protection. Sites also need compatible definitions and validation before combining evidence.

High-performance computing gains value through parallel work, but speed-up claims require the same task, hardware, precision, energy and accuracy boundaries. The chapter develops distributed-systems governance. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Draw data, model and permission flows across participating organisations. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Decentralisation does not automatically deliver privacy, fairness or equivalence.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC combines behavioural science, psychology, cognitive science and AI to model choices and interactions. Human data can be sensitive, context-dependent and altered by the intervention itself. Consent, representative sampling, construct validity and human review are central to responsible use.

Physics-based and data-driven models can complement one another when constraints, training evidence, uncertainty and failure cases remain inspectable. The chapter develops human-centred research. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a construct-to-measure map and test alternative explanations. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A population pattern must not be used to label an individual beyond evidence.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

IHPC-style teams include mathematics, physics, engineering, computer science, materials, chemistry, behavioural science and domain specialists. Students can reproduce a small simulation, profile it and compare it with observations. Current institutions control courses, access and professional roles.

Human-centred modelling needs consent, representative evidence and safeguards because a useful population pattern may still fail or harm an individual. The chapter develops computational-science career preparation. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build a small model with a simple baseline, validation plot and README. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A school simulation does not authorise operational or safety decisions.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Archive the question and intended use, equations, assumptions, parameter sources, data, code commit, environment, hardware, precision, verification, calibration, validation, uncertainty, human review and released decision. This lineage separates computational speed from scientific credibility.

Computational science turns equations, algorithms and data into testable representations of physical or human systems; validation decides where those representations can be trusted. The chapter develops evidence governance. Keep the question, input provenance, method version, control or baseline, uncertainty, review and next decision in one evidence receipt. That record shows what the research can support today and what still needs measurement, replication or governed approval. IHPC and its authorised collaborators control current projects, software and released results; domain owners, public agencies and regulators control operational decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a versioned system-to-code-to-validation-to-decision lineage. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: The latest authorised model, domain evidence and governance control use.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter