The networks that power modern civilization have run for decades on 1970s control logic, under-observed, under-optimized, and increasingly unequal to the demands of the energy transition. Our client was founded to change that.
They have already built a composite, physics-grounded causal world model: six co-trained inference engines spanning physics-informed GNNs, causal time-series, topology discovery, federated training, edge inference, and techno-economic optimization. It is in production at 30 distribution system operators across 14 countries, covering 22 million connection points. Several of the world’s largest industrial automation and grid-technology vendors integrate it into their platforms, and it runs on the live grids of multiple tier-1 European utilities today.
They are now building the next layer: a large-scale pre-trained foundation model for flow networks, trained on tens of billions of physics-consistent network states and governed by hard conservation-law constraints that no language model will ever learn from tokens. This is not applied AI. It is a new model class.
DeepRec.ai is partnering with the company to assemble a small, exceptional team to build it. The data is real, exclusive, and unglamorous. The physics is non-negotiable. The impact is continental.
The problem you will work on Distribution grids are among the most complex dynamical systems on Earth: millions of nodes, time-varying topology, hard physical constraints, and almost no labeled ground truth. The state of the art is classical SCADA with a thin ML veneer.
Our client is replacing it with a large-scale pre-trained foundation model, trained on synthetic and real network states, governed by Kirchhoff constraints as a hard loss term, and fine-tuned on operator-specific topologies via federated learning.
Stage 1 pre-training target: 10¹? Newton-Raphson power-flow solutions across 50,000 distribution topologies. Stage 2: cross-network generalization to gas, heat, and water flow networks. Same architecture, different conservation laws.
What you will do
- Own end-to-end pre-training of the physics-informed GNN foundation model: data pipeline design, masked pre-training objective, distributed training infrastructure, and evaluation harness.
- Characterize scaling laws for physics-informed pre-training: data efficiency vs. compute trade-offs, emergence of physical consistency, and OOD generalization across unseen topologies.
- Design the pre-training corpus: synthetic topology generation, power-flow simulation at scale, and augmentation strategies that preserve physical validity.
- Lead the foundation-model preprint: own the architecture and pre-training sections, targeting a top-tier venue (NeurIPS, ICLR, ICML) or arXiv first.
- Interface with the causal world-model team on physics-informed loss formulation, and with the federated training team on privacy-preserving pre-training across operator estates.
- Represent our client externally at frontier AI venues. We expect this person to be a recognizable scientific voice for the model class being defined.
- PhD in machine learning, computer science, or computational physics from a leading research institution (e.g. ETH Zurich, Cambridge, Oxford, TU Munich, EPFL, UCL, ENS, or equivalent).
- 3 to 6 years of post-PhD experience at a frontier AI lab or leading academic group (e.g. DeepMind, Meta FAIR, Mistral, EleutherAI, Stability AI, Kyutai, Aleph Alpha, Max Planck MIS, IDSIA, ELLIS-network member labs, or equivalent).
- First-author publications at NeurIPS, ICLR, or ICML on large-scale pre-training, masked modeling, GNN expressivity or scaling, or physics-informed deep learning.
- Hands-on experience training models at >1B parameter scale with distributed GPU/TPU infrastructure (PyTorch DDP/FSDP, JAX, or equivalent).
- Desirable: prior work at the intersection of graph neural networks and physical simulations, including molecular dynamics, fluid dynamics, power systems, or any PDE-governed network system.
- Desirable: experience with physics-informed neural networks (PINNs), neural operators (FNO, DeepONet), or Hamiltonian / Lagrangian networks.
- A genuinely unsolved research problem at the intersection of physics, ML, and critical infrastructure, with exclusive access to real production data from 30 grid operators.
- First-principles technical latitude: you define the pre-training objective, the architecture choices, and the evaluation methodology, subject to hard physical constraints, not product-manager preference.
- A small, senior team. You will work directly with world-leading researchers in physics-informed ML and graph-based power systems AI.
- Competitive compensation benchmarked to tier-1 European AI labs, with meaningful equity in a company with €4M committed capital and growing ARR.
- Publication and conference travel fully supported.
