Skip to content

Expertise

Research frontiers

Six frontiers. Three are developed in detail below; three are stated briefly. Each publishes only with at least two evidence links — project plus publication, simulator or benchmark.

01

Foundation models that learn physics from physics

We are building scientific foundation models trained to reason about simulation and to write the language our solvers execute.

  • Why

    A general language model has read about physics; it has not run any. Simulation generates unlimited correctly-labelled data on demand, and every answer can be checked against a solver — so the reward signal is verifiable rather than approximated.

  • Approach

    FPS-SciFM, with variants on Granite 4.1-8B-instruct and other architectures, trained by reinforcement learning from verifiable rewards (RLVR). FLUX is tokenizer-optimised and our tools curate training files automatically, so the corpus grows from our own simulation output.

  • State

    In continuing development. FPS-SciFM v1.0 model launched on SimuPort.

02

Agents that run simulations, not conversations about them

A self-improving agentic framework, deployed on SimuPort. Describe an engineering problem and it returns a solved, optimised model.

  • Why

    The agents write FLUX, the same language our solvers execute. Every generated model is a complete, readable definition — open it, change one term, re-run it, and anyone gets the same result. Reproducible, not one-off.

  • What it does

    Builds geometry and mesh, applies boundary conditions, runs and monitors the solve, batches runs, trains surrogates and optimises — iterating until the design meets your specification.

  • In parallel

    Separate agents mesh, check feasibility before and after each run, and couple solvers. Geometry comes from conversation, CAD, or an image our recognition turns into a 3D model or pipe network.

  • State

    Deployed on SimuPort today.

03

Runtime as a design constraint

Thousands of runs are only feasible if one run is cheap. Solver speed is what makes optimisation and surrogate training something a project can actually do.

  • Why

    In ALABAMA, RESTORE and WeldGalaxy, thousands of melt-pool and welding simulations trained the surrogates that optimisation then ran against. That workflow does not exist for a team whose single run takes a day.

  • Where it comes from

    In-house sparse-matrix preconditioned conjugate-gradient solvers, adaptive multi-resolution topology optimisation, optimised large eddy simulation, and a cluster we own and tune ourselves.

  • Measured

    published benchmark

04

Also advancing

  • Coupled multi-physics and multi-scale methods

    Simulation chains spanning density functional theory, thermodynamics, microstructure, continuum flow and mechanics, with surrogate stages replacing full physics where accuracy allows.

  • Uncertainty-aware optimisation

    Optimisation that accounts for variability in material properties and process parameters, so an optimised design still performs when recycled feedstock and real manufacturing tolerances are involved.

  • Open and interactive publication of computational research

    Publishing results as runnable, versioned, citable artefacts rather than static figures.