Domain-Specific Languages
Active library methods decouple scientific models from low-level backend complexity, supporting long-term maintainability and performance portability.
Research Portfolio
Algorithmic Abstractions for Heterogeneous HPC
Our central mission is to let domain scientists express numerical intent at a high level while generated code reaches the full capability of modern CPUs, GPUs, and distributed clusters.
Research Pillars
Production-Scale Applications
We validate methods in real applications including industrial CFD, large-scale microsimulation, and MRI analytics where computational efficiency materially changes project outcomes.
Scientific Reliability
Our work addresses the full lifecycle of scientific software quality: correctness, reproducibility, numerical precision, and platform-to-platform comparability.
OP2: Unstructured Mesh eDSL
OP2 models computation with sets, datasets, and mappings, enabling backend-specific optimization without rewriting scientific kernels.
| Backend Architecture | Programming Model | Optimization Mechanism |
|---|---|---|
| Multi-core CPU | OpenMP, SIMD Vectorization | Loop vectorization, cache blocking |
| NVIDIA GPU | CUDA | Memory coalescing, atomic operations |
| AMD GPU | HIP / SYCL | Performance portability via abstraction layers |
| Distributed Clusters | MPI | Automatic partition-based communication |
Industrial Validation
OP2 is a core acceleration vehicle for HYDRA, used by Rolls-Royce in gas turbine design. This demonstrates robust transfer from research prototypes to production-relevant workflows.
Academic Deployments
OP2 is also used in VOLNA (tsunami simulation), BASEMENT (river morphology modeling), and additional finite-volume / finite-element projects.
OPS: Structured Multi-block DSL
OPS focuses on stencil-heavy structured-grid workloads where data movement dominates runtime. Runtime loop tiling and backend code generation reduce memory-wall bottlenecks.
Performance Strategy
Automatic parallelization and backend specialization enable near hand-tuned performance in bandwidth-bound applications.
Representative Usage
OPS supports projects such as OpenSBLI for shock-boundary-layer interaction and SENGA2 for CFD and combustion research.
Interdisciplinary Case Studies
PanSim and COVID-19 Policy Analysis
Agent-based microsimulations enabled broad policy search across masking, mobility, and school closure strategies, with uncertainty-aware scenario exploration.
Virtual Certification for Aerospace
Collaborative CFD work contributes to reliable design workflows for next-generation gas turbine engines using performance-portable simulation pipelines.
Neuroimaging and Whole-Cell Modeling
GPU acceleration for diffusion MRI and drug-effect simulation enables data scales inaccessible through conventional single-node processing.
Technical Robustness and Trust
Bitwise Reproducibility
We develop graph-coloring based parallel and distributed algorithms that preserve deterministic floating-point behavior across decomposition strategies and process layouts.
Mixed-Precision Computing
Custom precision assignments (for example half/single/double combinations) improve performance and energy efficiency while maintaining required numerical fidelity.
Collaborative Network
Our research network includes world-class institutions such as the University of Oxford, Imperial College London, and the University of Warwick, with notable collaborators including Mike Giles, Paul Kelly, and Gihan Mudalige.