Applied Scientist - Scientific Modeling and Experiment Selection

The Role

A model predicts a material will work. We run the experiment. It doesn’t.

Was the model wrong? Is a variable missing? Are we outside its valid regime? Was the experiment noisy? What should we run next to find out?

You will make our scientific agent better at deciding which experiment to run next. You will build and adapt predictive physics, chemistry, and materials models, test them against laboratory evidence, investigate failures, and turn what you learn into better decisions. You will report to the Principal AI Scientist, Agentic.

The data is sparse and noisy. Experiments cost real time and money. We value getting ideas into meaningful tests quickly and learning from the results.

Campaign scientists run the physical experiments. ML engineers build the agent and infrastructure. You own the scientific modeling and experiment-selection logic between them.

The loop:

model → prediction → experiment → evidence → update → next experiment

What You'll Work On

• Build models that change experimental decisions. Select, develop, or adapt predictive scientific models for specific campaign tasks, starting with clear assumptions and simple baselines.
• Calibrate models for their intended use. Assess predictions and uncertainty against reference data. Establish where a model is reliable, where it extrapolates, and when its outputs need additional verification.
• Automate model-improvement experiments. Build reproducible workflows that compare model variants, parameters, calibration methods, and data choices within defined scientific constraints and compute budgets.
• Validate before agent integration. Evaluate candidate improvements on held-out data. Define acceptance criteria for predictive performance, uncertainty, robustness, and computational cost.
• Decide what experiment should come next. Use predictions, uncertainty, evidence, constraints, and cost to select promising or informative experiments.
• Investigate disagreements between models and the laboratory. Identify competing explanations and design tests that distinguish model limitations from missing variables, measurement noise, or execution problems.
• Turn scientific models into usable agent tools. Work with engineers to make inputs, units, assumptions, outputs, and failure conditions explicit. Define when the agent should query a model, seek additional evidence, or escalate to a scientist.
• Learn from every outcome. Preserve and interpret positive, negative, and inconclusive results, and use them to guide model updates and subsequent decisions.

What We're Looking For

We care most about whether you can independently move through:

problem → hypothesis → model → code → experiment → evidence → next decision

Required:

• Deep quantitative scientific training and 4+ years of hands-on experience, including doctoral research, adapting, or validating predictive physics, chemistry, or materials models against experimental evidence.
• Strong Python and scientific-computing skills, with the ability to turn an idea into a working model, reproducible experiment, or automated comparison workflow.
• Experience with Bayesian optimization, active learning, or experimental design, and an understanding of how predictions and uncertainty inform experiment selection.
• Sound judgment about small datasets, noisy measurements, model calibration, extrapolation, and conflicting evidence.
• The ability to establish appropriate baselines, keep development and validation separate, and distinguish a promising result from an unsupported conclusion.
• Clear communication with scientists and engineers, including explaining what a model can support, what remains uncertain, and what evidence is needed next.
• A PhD in Physics, Materials Science, Chemistry, Chemical Engineering, Applied Mathematics, Scientific ML, or a related field is strongly preferred. Equivalent industrial research experience is welcome.

You should design tests that challenge your own explanations, revise or abandon an approach when the evidence warrants it, and recognize when the evidence is still inconclusive.

Nice to have:

• Autonomous experimentation
• Lab automation
• Scientific workflow orchestration
• LLM agents

We do not expect all of these.

What Success Looks Like

Month 1: Take a real campaign problem. Establish the modeling baseline and evaluation criteria. Identify a consequential failure or uncertainty, and launch the first controlled test.

By Day 90: Deliver one quantitative approach evaluated against available laboratory evidence, an explicit policy for how it changes experiment selection, and a controlled comparison showing what improved, what did not, and what remains uncertain.

The model, configuration, data, and results should be traceable so someone else can reproduce the comparison. Where the evidence supports integration, work with engineers to make the approach available to the agent.

If it worked, quantify the improvement. If it failed, explain what you learned and decide what to test next.

Radical AI is an equal opportunity employer. We do not discriminate on the basis of race, color, ancestry, national origin, religion, sex, age, sexual orientation, gender identity and expression, marital status, disability, or veteran status.