ЕвропаMachine Learning Engineer — Physics AI
1 мес. назад
Европа
machine learningdata compressiongpu utilizationscientific computing
Hiring a Machine Learning Engineer to work on Physics AI and data compression technologies at Zibra AI, focusing on benchmarking and improving AI models for scientific computing and simulation.
Обязанности
- Benchmark our compression technology across a wide range of Physics AI architectures and datasets .
- Run large-scale experiments for CFD, turbulence, weather, engineering, and other scientific ML workloads.
- Measure the impact of compression on: model convergence and final accuracy;
- training throughput;
- GPU utilization;
- CPU and data-loading overhead;
- storage and network requirements.
- Compare compressed-data training against conventional pipelines and alternative compression methods.
- Research training directly in compressed or partially decoded representations .
- Explore compression-aware sampling, augmentation, tokenization, and model architectures.
- Design rigorous, reproducible benchmark methodology.
- Integrate compressed datasets into PyTorch and distributed training workflows.
- Turn experimental results into product recommendations and research directions.
- Write technical reports, benchmark publications, blog posts, and academic papers .
- Collaborate with external research groups and industrial partners on joint evaluations.
Требования
- Hands-on experience with Physics AI / scientific machine learning is required.
- Experience training models on simulation or physical-science datasets.
- Strong practical experience with PyTorch and modern deep-learning workflows.
- Familiarity with architectures such as: neural operators;
- mesh GNNs;
- transformers for physical systems;
- surrogate models;
- foundation models for science;
- PINNs or related methods.
- Experience with large 3D/4D datasets such as volumetric grids, meshes, point clouds, or spatiotemporal fields.
- Good understanding of GPU training performance, data loaders, profiling, and distributed training.
- Strong experimental methodology and ability to design controlled benchmarks.
- Ability to analyze how numerical approximation and preprocessing affect model quality.
- Strong Python and scientific-computing skills.
- Ability to communicate research results clearly in written technical form.
Будет плюсом
- Experience with PhysicsNeMo or similar scientific ML frameworks.
- Background in CFD, FEA, climate, turbulence, combustion, or computational physics.
- Knowledge of lossy compression, quantization, numerical error analysis, or signal processing.
- Multi-GPU or multi-node training experience.
- Previous academic publications in ML, scientific computing, compression, or related fields.
- If this role sounds like a fit — we’d love to hear from you! Just send over your CV and anything else you’d like us to consider.
- We’ll review everything within five working days , and if your background matches what we’re looking for, we’ll get in touch to set up a call and get to know each other better.
Другое
- We are , a Ukrainian recruiting company focused on hiring talent and building teams for tech startups worldwide. People make a difference in the big game, and we may help find the right ones.
- Currently, we are looking for a Machine Learning Engineer — Physics AI for Zibra AI .
- Zibra AI is a deep-tech company building advanced technologies for working with large-scale 3D data. The team has a strong background in computer graphics and data compression and is now expanding its technology into industrial simulation and Physics AI.
- The company is developing a new data infrastructure layer that makes massive scientific and simulation datasets significantly easier to store, transfer, visualize, and use for AI training.
- You will work at the intersection of Physics AI, scientific computing, ML systems, and data compression. A major part of the role is to benchmark our codec across different model architectures, study how compression affects accuracy and training efficiency, and explore new approaches to training directly in compressed representations.