Physics-Informed Machine Learning Engineer

Svitla Systems Inc. is looking for a Physics-Informed Machine Learning Engineer for a full-time position (40 hours per week) in Ukraine. Our client is a technology startup.

The team is building a Physics-Informed Foundational Model to understand GPU and compute health. They derive physics-grounded stress signals: effective-stress proxies, semiconductor degradation estimates, and dynamical mathematical features, and use them to assess hardware health over time. The feature pipeline runs end-to-end. You'll build the model and the fusion layer on top of it. You will own the modeling: designing the fusion layer (how physics-based and dynamical features combine into a coherent health signal) and the temporal modeling layer (a physics-informed model, PINN-style, where empirical stress signals drive part of the loss and a physics-based degradation model informs another part). The exact formulation of the physics term is still evolving; you'll be involved in shaping it. You'll work as part of a small, technical team alongside the founder and other domain experts.

• Experience in building and training physics-informed models - a physics-based term in the loss function of a real project (PINN, physics-regularized NN, or equivalent).
• Strong understanding of time-series/sequence modeling (LSTM, temporal CNN, transformers, or state-space models) on sensor or telemetry data.
• Understanding of parameter calibration / inverse problem: fitting mechanistic model parameters to noisy observational data (Bayesian calibration, MLE, or optimization-based).
• Expert knowledge of Python scientific stack (Pandas, NumPy, scikit-learn, PyTorch or JAX) and be comfortable owning a data pipeline end to end, including data-quality investigation.
• Expertise in reading and reasoning about physics/reliability equations governing degradation; you don't need to derive them, but they can't be a black box.
• Experience in reliability engineering/PHM (prognostics and health management) background: RUL estimation, degradation modeling, accelerated-life testing.
• Exposure to semiconductor or hardware degradation physics at a "read the literature critically" level.
• Familiarity with nonlinear dynamics/recurrence or dynamical-systems features (e.g., RQA or comparable techniques).
• Familiarity with hardware/datacenter telemetry or fleet analytics.
• Experience working in small teams alongside domain scientists/mathematicians; comfortable turning research feedback into production code.
• Build the temporal model: design and train a physics-informed sequence model (e.g., LSTM or similar temporal architecture) for degradation and health prediction, incorporating a physics-based loss term alongside the data-driven loss.
• Design the fusion layer: define how physics-based stress features, dynamical/mathematical features, and other signals combine into model inputs and a defensible health score, replacing today's simple hand-set weighting.
• Calibrate the physics-informed components: the stress-proxy parameters are currently engineering priors. You'll help design and execute calibration strategies against whatever outcome labels are available.
• Harden the feature pipeline: the pipeline is Python/Pandas over time-aligned multi-sensor telemetry; you'll extend and maintain it (feature audits, label engineering, data-quality gates) as modeling needs dictate.
• Write clear analysis docs and defend modeling choices to technical stakeholders and clients.