Materials & Process Scientist

Atmospheric AI; Greater London; Hybrid; GBP 70,000 - 110,000; Full time

Job summary

Atmospheric AI Limited in London is seeking a strong modelling specialist to build physics-based transformation and conversion models for materials economics and carbon accounting. You will enforce mass and energy balances and validate models against plant evidence, licensor data and literature.

You will work on canonical material tables and interfaces with Python-based tooling, Databricks or MLflow, and collaborate with teams to ensure robust, production-ready models.

Qualifications

• Strong chemical and process-engineering fundamentals required.
• Experience with mass and energy balances and thermodynamics.
• Ability to design validation tests and set tolerances.

Responsibilities

• Build transformation models with energy and mass balance constraints.
• Develop conversion models for materials processing (extrusion, molding, etc.).
• Create a durable canonical materials table and ensure data integrity.
• Implement automated validation tests against model outputs.
• Work across metals, paper, fibres or new chemical systems.
• Contribute to a robust data-modelling framework using Python.

Skills

Mass & energy balances

Thermodynamics

Reaction engineering

Process economics

Unit operations

Materials knowledge

Validation mindset

Python modelling

Modular code

Canonical tables

Education

Master's or PhD in chemical, process or materials engineering or polymer science

Tools

Python

PostgreSQL

Dagster

AWS

Databricks

MLflow

Job description

Everything physical is made from raw materials: polymers, chemicals, metals and minerals, a $16T economy that software has barely modelled. Atmospheric AI computes what those materials should cost, what carbon they carry and where their supply is exposed, from the physics of how they're made, so manufacturers can decide and negotiate from ground truth.

The problem

Our cost, carbon and substitution answers are computed from process and material models. A buyer asking whether one grade can replace another, or what a converter's energy bill does to its price, is asking those models a question. They have to respect mass and energy balances, connect grades to the technologies that make them, and say plainly where they stop being valid. Physics has the final say, and you'll have the authority to block a release when the evidence doesn't hold.

What you'll own

• Transformation models. Industrial processes with credible mass and energy balances, yields, utilities and by-products, within thermodynamic and reaction-engineering constraints.
• Conversion models. How materials become intermediate and finished products: extrusion, moulding, film and fibre, forming and finishing.
• The material layer. Grade comparison, substitutability, and which technologies and units could have made a given grade.
• Canonical scientific tables. Shared identifiers, classifications and hierarchies for materials, chemicals, properties, grades and processes, so the whole product speaks one scientific language.
• Validation as tests. Benchmarks against simulation, licensor, literature and plant evidence, with physical constraints turned into automated tests.
• The quality bar. Models with stated coverage, uncertainty and limits. Nothing ships that fails the physics.

What we're looking for

• Strong chemical and process-engineering fundamentals: mass and energy balances, thermodynamics, reaction engineering, separations, process economics, unit operations.
• Materials knowledge. How composition, structure and production technology set grade properties and substitutability.
• A validation mindset. You design tests that could prove your model wrong, set tolerances and withhold approval when the evidence is thin.
• Strong Python modelling. Modular, tested, reproducible code someone else can review and maintain, and the data-modelling sense to keep canonical tables durable.
• Depth in one family, reach beyond it. You can carry your reasoning into metals, paper, fibres or new chemical systems. Walk we…

Nice signals: Aspen Plus, HYSYS or comparable simulation; technical datasheets, grade databases or licensor data; taking a scientific model into production yourself; Databricks and MLflow or equivalent; a Master's or PhD in chemical, process or materials engineering or polymer science.

Stack: Python, FastAPI, Postgres, Dagster, AWS. A signal, never a requirement.

The take-home: we send you a small transformation model with one physical error in it. Find it, write the tests that would have caught it, and propose the fix.