PRISM-alpha has been open-sourced, and an ESA EXPRO+ proposal (Ref: ESA AO/1-13197/25/FR/LCF) has been submitted under the FIRST! Simulation & Intelligence Technologies call. The proposal targets a 12-month de-risking timeline with a four-partner consortium.
What PRISM is
PRISM (Platform for Research in Intelligent Synthesis of Materials) is an autonomous discovery engine that replaces the traditional multi-year materials R&D cycle with AI-driven computational screening and robotic synthesis. The initial target: refractory high-entropy alloys for oxygen-rich preburners in liquid rocket engines, replacing legacy materials like Monel K500.
The system runs a dual-loop workflow. A computational inner loop screens thousands of candidate alloys per hour. A physical outer loop validates the top picks through robotic synthesis and hot-gas testing. The two loops feed each other — experimental results retrain the models, and the models propose new candidates that the experiments wouldn't have found.
Architecture
Four modules work together:
The Evolver is the orchestrating reasoning model. It uses Agentic Context Engineering to build and refine a "playbook" of discovery strategies, balancing exploration (try new composition spaces) and exploitation (refine promising candidates) through a Generator, Reflector, and Curator — without suffering context collapse over long campaigns.
The Mutator Fleet consists of specialized AI sub-agents that propose candidate modifications. A Composition Mutator handles elemental substitutions. A Process Mutator optimizes synthesis protocols. A Thermo Mutator queries thermodynamic databases and checks phase stability. Each mutator operates independently, and the Evolver decides which mutations to accept.
The Evaluator is a multi-tiered screening pipeline that provides the fitness score. Fast surrogate models (MACE-MH-1 equivariant graph neural network) and ab initio DFT/CALPHAD checks triage candidates computationally, routing only the best to physical A-Lab robotic synthesis for ground-truth data. This tiered approach keeps the cost per candidate low enough to evaluate thousands per run.
The Materials Knowledge Graph is the system's active memory. It ingests heterogeneous data from scientific literature, patents, and experiments. The Evolver queries it to break out of search stagnation and formulate new hypotheses based on existing knowledge rather than random exploration.
The consortium
The ESA EXPRO+ proposal is structured around four partners:
- Bimo Tech (Prime) — AI architecture, PRISM digital design, surrogate model development, LPBF synthesis of down-selected alloys
- Fraunhofer IAPT — AI-based Design for Manufacturability scoring, R&D LPBF process calibration, defect evaluation
- ArianeGroup — Industrial end-user requirements, physical validation via ERBURIGK hot-gas oxidising facility, benchmarking against Monel K500
- amsight — Data integration, structuring manufacturing and qualification data into machine-readable formats for model retraining
Where it stands
An earlier ESA OSIP market research study on the concept was positively evaluated. Active data pipelines with NCBJ and IPPT PAN in Poland are in place for scaling and subscale testing. The pending Phase 1 kick-off focuses on a Cold Start bootstrap — using transfer learning to get the system running from near-zero domain data, so it doesn't need a massive training set before it can start discovering.
PRISM's fitness function is defined by S.P.A.R.K.'s alloy targets. What SPARK identifies as the engineering requirements, PRISM uses as its optimization objective.