Nature Portfolio · July 2021 Teaching solid mechanics to artificial intelligence — a fast solver for heterogeneous materials
Research
Research that holds up in peer review.
We build AI systems for the Mittelstand. We test our methods where scrutiny is hardest: in peer-reviewed journals, with the Max Planck Institute and with university researchers. The same evidence standard goes into every system we deliver.
- 8,300×
- faster than a conventional solver, published with the Max Planck Institute
- 168
- citations of the article in the literature, as of November 2026
- 4
- collaborations with researchers since 2020, in Germany, Argentina and India
Nature Portfolio · July 2021 With the Max Planck Institute
We taught solid mechanics to an AI.
In July 2021, DeepMetis and the Max-Planck-Institut für Eisenforschung (MPIE) in Düsseldorf published a deep-learning solver. It predicts local stresses in heterogeneous materials up to 8,300 times faster than a conventional numerical solver.
The problem
How complex materials respond to load decides whether cars, bridges and machines are safe. Conventional solvers are accurate but iterative. A single simulation of a non-linear material takes minutes to hours.
Our approach
We trained a modified U-Net, a convolutional neural network, on 1,000 full-field simulations computed with MPIE's open-source DAMASK software. Once trained, the network returns the full stress field in a single forward pass. No iteration, including for geometries it never saw in training.
Our role
DeepMetis co-developed the concept, designed the neural network and carried out its training. Dr. Jaber Rezaei Mianroodi (MPIE) generated the simulation data. Prof. Dierk Raabe, MPIE director, supervised the work.
Why it matters for industry
Fast surrogate models make simulation-heavy tasks practical: design and topology optimisation, early identification of stress hot spots, and faster starting points for conventional solvers.
Published open access in npj Computational Materials (Nature Portfolio) and cited 168 times to date. MPIE now operates as the Max-Planck-Institut für Nachhaltige Materialien.
What we are researching right now.
Three ongoing projects on questions that feed straight into our client work: how LLM systems stay reliable in production, what software work really consists of, and when a company is ready for AI.
Quality assurance for LLM systems
So that AI doesn't quietly break after go-live.
Institutions involved 
Uka Tarsadia University Bardoli, India
Uka Tarsadia University Bardoli, India With Sapan Naik, assistant professor at Uka Tarsadia University (Bardoli, India), and two of his students. Joint research paper, started in January 2025.
LLM systems are unstable. Changing a prompt, swapping the base model or adding a tool fixes one behaviour and silently breaks another.
The framework comes from our research and development team. Every LLM step becomes an encapsulated component with fixed input and output schemas. This makes quality drift visible before customers notice it.
The framework is model-agnostic. It combines simple checks such as schema validity and required strings with LLM-based grading of factuality, similarity and criteria. We use the same principle in our client systems today.
How the test pool grows
- A bug, a bad response or a new feature Every case is recorded, not just fixed.
- Becomes a test case With fixed input and expected output, in a growing pool.
- Every change runs against all cases The new ones and all earlier ones, whether it's a prompt, a model or a tool.
- Ships only when all of them pass What worked before demonstrably keeps working.
Our own research
Two questions every client asks us.
In 2026 we began publishing our own research: what does a project really consist of? And is an organisation ready for AI at all? Both papers state openly what is already evidenced and what is still a hypothesis.
WIT
An Intent-Centered Framework for Classifying Software Work Items
- Question
- Why does a ticket exist, and how should work be classified across tools such as Jira, Azure DevOps and GitHub?
- Method and evidence
- Case study of 183 tickets, 50 meeting transcripts and project documents. Six intent types, inter-rater agreement κ = 0.72. Piloted on 244 open-source GitHub issues.
AIRO
An AI Readiness Framework for SMB and Mid-Market Enterprises
- Question
- Which kind of AI can a small or mid-sized company realistically absorb, and with what return?
- Method and evidence
- Design-science framework with 4 dimensions and 25 sub-dimensions. Scoping review of about 120 sources, 37 scorable DACH cases. Delphi study and fsQCA validation planned.
The quality of connections between a company's systems predicts AI success better than the sophistication of any single system.
It reflects what we see in practice: most business AI reads from one system and writes to another.
More collaborations
Research since our founding year.
With a TU Berlin researcher
Explainable AI for critical decisions
In our founding year we worked on explainable AI (XAI): methods that show why a model reached a decision, a precondition for using AI in medicine, insurance and public administration. Our research team developed “smart highlighting”. It marks the passages in patient files that drove a model's assessment, so clinical researchers can review pre-existing conditions faster and verify each decision.
On 17 December 2020 we hosted the online event “Explainable AI: Opening the Algorithmic Black Box for Critical Decision Making”. Speakers included Damien Foucard, machine-learning researcher at TU Berlin, on XAI-driven summarisation of dynamic data, alongside our head of research and Tarek R. Besold of neurocat. The exchange with TU Berlin continued into 2021 through joint technical workshops.
Research group “Seguridad en Democracia”, Argentina
AI tooling for public-interest research
The group at Universidad Nacional del Sur, led by Dr. Tobías J. Schleider, studies the links between new technologies and violence prevention. We gave it free access to our LLM agent platform. It used generative AI to draft, review and translate papers, summarise literature, analyse data and check for plagiarism, always as groundwork for public-policy design.
The collaboration ran both ways. The group's structured feedback, for example on conversation history, exporting results and translating longer texts, went directly into our product backlog.
The partnership refers to the research group “Seguridad en Democracia”, not to Universidad Nacional del Sur as an institution.
In your project
Research is not a side project for us. It decides what we build.
-
Literature reviews before architecture decisions
Before we settle on an architecture, we check what research and practice already know. A 2026 review of work on multi-agent systems for software engineering now shapes how we structure our own delivery.
-
Platform research
We document the data models of the ERP systems we replace or integrate, from vendor sources. Among them Microsoft Dynamics 365, Sage, ERPNext, Odoo and Oracle NetSuite.
-
Technology surveillance
A recurring internal brief tracks new models and open-source tools and rates their maturity before we use them for clients.
-
Verified numbers only
Every statistic we use is traced to its primary source and graded as verified, partially verified or unverified. Unverified figures do not reach client documents.
Your system deserves the same standard.
Show us your process. We'll tell you honestly where AI helps, where it doesn't, and what that is based on.