Materials Informatics Scientist Job Description

A Materials Informatics Scientist uses data, statistics and machine learning to decide which materials a research team should make and test next. The candidates might be battery cathodes, high-temperature alloys, polymer formulations, catalysts or semiconductor films. Instead of running long trial-and-error campaigns at the bench, the informatics scientist assembles what is already known, trains models on it, and hands the lab a ranked shortlist or an experiment plan. Success is judged in the lab: fewer failed syntheses, faster convergence on a target property, and a dataset that is more useful at the end of a project than at the start.

Where the data comes from

The raw material is a mix of computed and measured data. Computed data comes from density functional theory (DFT) runs and from open databases such as the Materials Project, which its director said in an April 2025 DOE Office of Science interview includes information on 150,000 materials and has more than half a million registered users. Measured data comes from an employer's own instruments: X-ray diffraction patterns, electron micrographs, mechanical tests, conductivity and battery cycling data, scattered across spreadsheets, electronic lab notebooks and instrument PCs. The first deliverable on a new project can be nothing more glamorous than clean, versioned tables with consistent compositions, units and processing histories.

The modeling work

With data in hand, the scientist builds features from compositions, crystal-structure graphs and microstructure images, then trains models that predict properties such as formation energy, band gap, ionic conductivity or yield strength. The everyday toolkit is Python with pymatgen and matminer for materials data, scikit-learn for classical models, PyTorch for graph neural networks, and Gaussian-process or Bayesian optimization libraries for picking the next experiment. Generic machine learning code does not drop in cleanly. NIST researchers building open-source tools for autonomous experimentation note that in materials science "the data has significant physical meaning," and that algorithms must be adapted to account for those physical constraints. In practice that means respecting charge balance, phase stability and the limits of what a furnace or sputtering tool can actually do.

Closing the loop

The most automated version of the job wires models directly to robots. NIST describes closed-loop autonomous systems that "place machine learning in control of experiment design, execution, and analysis." In that setup the informatics scientist writes the decision logic (acquisition functions, stopping rules, safety limits), integrates instrument data streams, and audits what the system concluded before anyone publishes or scales it.

Who they work with and how the job differs

Day to day, the role works with synthesis chemists, process engineers, characterization staff, computational scientists running simulations, and software engineers who maintain shared infrastructure. It differs from a computational materials scientist, whose core output is a simulation study of a specific system: the informatics scientist's output is a model or decision system trained across many results, simulated and measured. It also differs from a bench materials scientist. The informatics scientist may never load a furnace, but owns the logic that decides what goes into it, and answers for it when the prediction misses.

There is no single credential for this title, so hiring managers look for two halves of a skill set: real materials science depth and working machine learning. Either half alone is a different job.

Education. A typical route is a PhD in materials science and engineering, chemistry, physics, chemical engineering or mechanical engineering, with a dissertation that used computation or data-driven methods. A master's degree paired with strong programming and a portfolio can fit applied industrial roles. Computer science and statistics graduates can enter too, but they need to show command of crystal structure, thermodynamics, phase diagrams and processing-structure-property relationships.

Graduate certificates. Some universities package the hybrid skill set into short academic credentials. NC State's Graduate Certificate in Materials Informatics is designed for graduate education "at the intersection of materials science, engineering, and data science," and its core course, MSE 723, emphasizes practical implementation of machine learning techniques on materials science problems. Texas A&M's Department of Materials Science and Engineering offers a certificate in Materials, Informatics and Design that requires completion of four courses (12 credit hours). These are university certificates, not licenses, and no professional certification is standard for the role.

Technical skills to build.

• Python fluency: pandas, NumPy, scikit-learn and a deep learning framework such as PyTorch
• Materials libraries such as pymatgen, ASE and matminer, plus the ability to read DFT outputs from codes like VASP or Quantum ESPRESSO
• Statistics: uncertainty quantification, design of experiments, Gaussian processes and Bayesian optimization
• Data engineering basics: SQL or MongoDB, Git, reproducible environments and job scheduling on high-performance computing clusters
• Working knowledge of characterization data, such as indexing an XRD pattern or segmenting a micrograph

Experience. As a rough guide, entry roles fit candidates with a relevant PhD or 2 to 4 years of applied work after a master's, and senior roles fit people with 5 or more years who have shipped models that changed what a lab actually made. Bring evidence: a public code repository, a published dataset, a paper in which a model's prediction was confirmed by synthesis, or an internal tool that other scientists adopted.

Working style. The job runs on translation. A strong informatics scientist can explain to a synthesis chemist why the model wants an odd composition, and says plainly when the model is extrapolating outside its training data. Patience with messy legacy data and the discipline to document every cleaning and modeling decision are part of the job description, not extras.

BLS does not track materials informatics as its own occupation, so the closest official numbers are for materials scientists. In its Occupational Outlook Handbook, BLS projects materials scientist employment to rise from 8,600 jobs in 2025 to 9,300 in 2035, and projects about 6,500 openings a year on average across the combined chemists and materials scientists group. BLS does not report informatics specialists separately within those counts. In the scientific research and development services industry, BLS reported a May 2025 median wage of $132,600 for materials scientists.

Federal programs aimed at AI-driven discovery

Two federal efforts shape where research money points. The Materials Genome Initiative's 2021 Strategic Plan called for challenges to promote adoption of its Materials Innovation Infrastructure, including autonomy, artificial intelligence and robotics, and the 2024 MGI Challenges named five problems informed by input from the MGI community. On Oct. 30, 2024, the CHIPS R&D Program announced the CARISSMA funding opportunity for AI and autonomous experimentation in sustainable semiconductor materials and processes.

The Genesis Mission, created by Executive Order 14363 on Nov. 24, 2025, put the Department of Energy in charge of a national platform connecting supercomputers, experimental facilities, AI systems and datasets. One of the focus areas on DOE's Genesis Mission page is "Discovering mission-ready materials," aimed at alloys, polymers and composites for defense and industry, and NSF issued a Dear Colleague Letter inviting proposals in support of the mission. Funding priorities can shift with each budget cycle, so read these as signals of where research money is aimed, not as hiring guarantees.

What AI has and has not done

The headline results show both promise and limits. In a 2023 Nature paper, researchers reported that their GNoME models found 2.2 million crystal structures stable with respect to the Materials Project, increasing the number of known stable crystals by almost an order of magnitude. Predicting stability is not the same as making something useful. A 2025 Materials Horizons study that benchmarked generative models against simple baselines found the baseline methods excel at proposing stable materials, while generative models outperform baselines at targeting properties such as band gaps near 3 eV; both approaches face challenges finding materials with exceptional properties.

That gap between a prediction and a working material is where this job lives: choosing which predictions deserve lab time, checking them against measurements, and feeding the results back into the next model.

Where the jobs are and where they lead

Employers include national laboratories, chemical and specialty-materials makers, battery and semiconductor companies, aerospace and metals producers, contract R&D firms, and startups that sell materials informatics software. Career paths lead to principal scientist, leader of an informatics or AI-for-R&D group, or technical product roles at software vendors. Senior roles put the person in charge of what a lab makes next, which takes modeling skill plus credibility with the experimentalists who have to run the work.

Dear Hiring Committee,

I am applying for the Materials Informatics Scientist position on your solid-state battery team. I finished my PhD in materials science last year, and my dissertation sat squarely between the glovebox and the GPU: I built property models for sulfide electrolytes and then helped the synthesis group test what those models recommended.

The part of that work I am proudest of was not the model itself. When I joined the project, our conductivity measurements lived in four spreadsheets with three different naming schemes for the same compositions. I spent my first semester building a single versioned dataset, tying every measurement to its precursor batch, annealing schedule and diffraction pattern. Only then did a graph neural network trained on our data and public DFT results become useful. I paired it with a Bayesian optimization loop, and over the following year the group used its suggestions to choose its synthesis targets. Two of those compositions became the basis of our best-performing cells.

I also learned where models break. Our first round of recommendations included compositions that were thermodynamically plausible but impossible to process in our furnaces. I rebuilt the search to include processing limits and an uncertainty threshold, and I started sitting in on synthesis meetings so I could hear why chemists rejected certain suggestions. That habit made the models better and made the chemists trust them.

I work in Python every day with pymatgen, scikit-learn and PyTorch, run jobs on our university cluster, and keep every analysis in a shared repository with documented data-cleaning steps so others can reproduce it.

Your team's plan to connect its new automated synthesis line to a decision model is exactly the problem I want to work on next. I would welcome the chance to discuss how I could help design that loop and keep it honest.

Sincerely, Maya Okafor