Scientific AI & HPC consulting
Is your simulation too slow?
I help engineering teams decide whether scientific machine learning can make a simulator fast enough—and build a validated prototype when it can.
My work covers neural operators, physics-informed methods, surrogate modeling, GPU optimisation, and high-performance computing for CFD, plasma, electromagnetic, heat-transfer, and multiphysics problems.
Tell me about your simulationSelected consulting and research results
Moldex3D simulation speedup: hours to seconds with less than 3% error.
Faster lower-hybrid-wave inference in a rectangular domain; 25.8× in TST-2 circular geometry.
NeurIPS 2024 ML4CFD Competition, with the highest ML score among participants.
The expensive, awkward, not-quite-production-ready ones.
- A numerical solver that takes too long to run.
- A parameter sweep that is too expensive to repeat.
- A surrogate that is fast but not yet trustworthy.
- A workflow that needs better use of GPUs or a compute cluster.
- A CFD, plasma, electromagnetic, heat-transfer, or multiphysics model that may benefit from operator learning.
Start with the real problem.
- 1
Send the solver, not a pitch deck.
Share the geometry, current runtime, available reference data, and what “fast enough” means for your team.
- 2
Check feasibility.
I assess whether a neural surrogate is appropriate—and say when classical HPC is the better answer.
- 3
Build and test a prototype.
The surrogate is benchmarked against reference data with an explicit error budget.
- 4
Hand it over.
Your team receives the model, code, benchmarks, integration guidance, and technical context needed to continue.
Small enough to test. Serious enough to trust.
Feasibility study · Prototype development · Technical consulting · Research collaboration
Start with an email.
Tell me what runs slowly and what a useful result would look like. I’ll reply with an honest first read.