AI Researcher: Multimodal Understanding

Mirror Physics•New York City, NY

1d•Hybrid

About The Position

The Company Mirror Physics is a New York-based AI company working on a new frontier in scientific simulation. We design intelligent systems that understand physics from first principles, providing critical acceleration for advanced technological R&D. Today, we’re building the world’s most capable AI platform for predicting experimental outcomes in chemistry and materials science, tightly coupled with reality via high-throughput experimental verification, to accelerate discovery in biotech, energy, manufacturing, and other domains. Backed by leading investors and scientific experts, we are expanding our research team at a pivotal moment in the field. The Opportunity Designing materials and chemicals to address society’s most pressing challenges will require joint reasoning across the domains of language, vision, and chemical structure. As Mirror’s lead of multimodal AI, you will architect and train multimodal generative models to understand, navigate, and design within the scope of physics, chemistry, and materials science, using high-quality observational data to fuel concept discovery at scale.

Requirements

Ph.D. or M.S./B.S. with equivalent research record in Computer Science, Materials Science, Chemistry, or related field with emphasis on machine learning
3+ years research experience with at least two modalities (language, vision, 3-D molecules/point clouds, graphs).
Fluency in Python plus modern ML stacks (PyTorch/JAX) and familiarity with distributed training tooling (CUDA, NCCL, Slurm/K8s/Ray).
Proven track record shipping large-scale generative or contrastive models (e.g., LLMs, diffusion, VAE, CLIP, NeRF, equivariant GNNs).
Strong publication or open-source track record in ML for physical sciences.
Excellent collaboration, communication, and team-working skills.
Deep commitment and passion for advancing science.

Nice To Haves

Familiarity with quantum chemistry, atomistic simulation, or chemistry/materials science
Familiarity with active learning, retrieval-augmented generation, or agentic workflows for scientific automation.
Prior work aligning multimodal models with scientific knowledge bases or ontologies.

Responsibilities

Design & build multimodal architectures fusing text, vision, and chemical structure.
Curate & align heterogeneous datasets from scientific literature, experimental data, and high-fidelity physical simulations.
Develop training pipelines for large-scale pre-training and instruction tuning on multi-GPU/TPU clusters
Create novel evaluations for cross-modal reasoning (text→structure, image→reaction pathway, etc.) and physical consistency.
Collaborate with AI, applied science, and engineering teams to build a product pipeline by integrating multimodal embeddings into search, reasoning agents, and generative design loops.
Engage with the AI-for-science community through publication and contributions at NeurIPS, ICML, ICLR, or other domain venues.