BIO

Dr Ado Farsi is a research fellow co-investigator at University College London and a visiting researcher at Imperial College London. He is also the co-founder of Tanuki Technologies.

His expertise lies in computational mechanics, particularly finite element (FEM) and discrete element (DEM) modelling, as well as machine learning (ML). He has served as an invited speaker and session organiser at international conferences including WCCM, ECCOMAS, USNCCM and ARMA.

Dr Farsi has extensive experience in solid and fluid mechanics, with a focus on Earth materials (e.g., rock and ice) and engineered materials (e.g., concrete and ceramics). His research spans applications from rock mechanics and fibre-reinforced concrete tunnels to the optimisation of catalyst supports for hydrogen production and the simulation of Arctic sea ice.

More information about his research is available at adofarsi.com.

PROPOSAL

This proposal will support Dr Ado Farsi as the named Research Technical Professional (RTP) based at Imperial College London (Department of Mathematics) as a member of the Firedrake team. The project will upskill Dr Farsi in AI for Science and Engineering through hands-on development and deployment of machine-learning operator learning capabilities within Firedrake’s diXerentiable finite element ecosystem. Dr Farsi will align primarily with the Explore AI stream (with elements of Excel AI), strengthening skills in operator learning, scientific ML, robust training workflows, and sustainable research software engineering.

The focus will be hybrid physics-based/ML simulation of physical systems arising in scientific and engineering contexts. The simulation of systems based on fundamental physical laws has revolutionised science and engineering. From weather forecasting to structural mechanics and medical imaging, simulation is everywhere. However, a key limitation of this approach is that real systems often have components for which we do not fully know or cannot represent the underlying physics. Combining physics-based simulation with data-driven ML operators oXers the opportunity for accurate simulation, prediction and design beyond the state of the art. More fundamentally, it enables the discovery of hitherto unknown operators by combining incomplete theory with observed data.

The key software challenge of the hybrid physics-based/ML is that the combined system must be trained on the data available. This involves diXerentiating (back-propagating) the combined system. Firedrake oXers a unique capability among advanced physics-based simulation frameworks: it employs the same diXerentiable programming abstraction used by the ML frameworks such as JAX and PyTorch.

Consequently, Firedrake simulations are fully diXerentiable. Further, Firedrake has been coupled with JAX and PyTorch and Dr Farsi has pioneered the hybrid approach, producing proof of concept results learning unknown constitutive laws in material simulation (Farsi et al. 2025).Dr Farsi is an established expert in physics-based structural modelling across the geosciences and engineering, who has been developing AI/ML skills through his work on operator discovery. This project will be an opportunity for him to gain a deep, theoretically sound basis in AI/ML techniques and to apply them through the development of advanced and robust hybrid simulation capabilities for the entire Firedrake community.

Developments will be merged into the Firedrake codebase and form part of the six-monthly Firedrake releases. This makes the capability available to the community, and ensures its sustainability as a part of Firedrake’s maintained releases which have been sustained for over a decade. Alongside technical developments, hybrid capabilities will be documented in Firedrake’s manual and demo suite, and demonstrated at training events over the year of the project.

Initial demonstrator applications will focus on Dr Farsi’s expertise in constitutive laws for structural models, but the resulting technology will be applicable across Firedrake’s diverse applications including for unresolved processes in the Gusto atmospheric code (Exeter and Met OXice), fusion reactor simulations (UKAEA and Los Alamos National Laboratory), battery chemistry (Lawrence Livermore National Laboratory) and many others.

Farsi, A., Bouziani, N., & Ham, D. A. (2025). Missing Physics Discovery through Fully DiXerentiable Finite Element-Based Machine Learning. arXiv preprint arXiv:2507.15787.