PRISM

AI-Native Autonomous Materials Discovery

MARC27 — ESA SPARK Prime Contractor | ITER Supplier

Quick Start • Architecture • Capabilities • License

---

PRISM is an AI-native platform for autonomous materials discovery. It combines large language models, multi-step agent orchestration, CALPHAD thermodynamics, ML property prediction, and federated data access into a single CLI that researchers can use to go from a hypothesis to a validated alloy candidate.

Built for the OSIP programme, PRISM targets refractory high-entropy alloys (RHEAs) for space propulsion but is domain-agnostic by design.

Architecture

PRISM implements a four-module closed-loop architecture inspired by biological evolution:

Module	Role	Current Implementation
Evolver (ACE)	Propose candidate compositions	Agent + GFlowNet (Phase G)
Mutator Fleet	Perturb and explore the design space	Skills + tool chaining
Evaluator	Three-tier validation (surrogate / CALPHAD / experiment)	ML predict + CALPHAD bridge + validation rules
MKG	Materials Knowledge Graph for memory and retrieval	Session memory + DataStore + scratchpad

The agent runs a Think-Act-Observe-Repeat (TAOR) loop with provider-agnostic LLM backends (Anthropic, OpenAI, OpenRouter), a tool registry, and skill orchestration.

Capabilities

Data Access

• 40+ databases via OPTIMADE federation (Materials Project, OQMD, COD, JARVIS, AFLOW...)
• Materials Project native API with formation energy, band gap, hull distance
• OMAT24 (Meta) via HuggingFace streaming
• Literature search (arXiv + Semantic Scholar)
• Patent search (Lens.org)
• Local data import (CSV, JSON, Parquet)

AI & ML

• Property prediction with auto-training (Random Forest, XGBoost, LightGBM)
• Feature importance and correlation analysis
• CALPHAD phase diagrams, equilibrium, and Gibbs energy (pycalphad)
• Simulation planning with auto-routing (CALPHAD vs DFT vs MD)

Agent Orchestration

• 10 skills: acquire, predict, visualize, report, select, discover, simulate, analyze phases, validate, review
• Plan-then-execute: agent proposes a plan, user approves before execution
• Tool consent: expensive operations require explicit approval
• Scratchpad: append-only execution log for reproducibility
• Feedback loops: validation and CALPHAD findings feed back into agent context

Infrastructure

• Two modes: interactive REPL (prism) and autonomous (prism run "goal")
• MARC27 managed LLM: /login for managed model access via MARC27 account
• MCP server: prism serve exposes tools and resources via FastMCP 3.x
• Plugin system: pip entry-points + ~/.prism/plugins/ local plugins
• Session memory: save, load, and resume conversations
• Reports: Markdown, HTML, and PDF with embedded charts

Quick Start

One-command install

curl -fsSL https://prism.marc27.com/install.sh | bash

pip install (from GitHub)

pip install "prism-platform[all] @ git+https://github.com/Darth-Hidious/PRISM.git"

From source

git clone https://github.com/Darth-Hidious/PRISM.git
cd PRISM
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[all,dev]"

Configure

prism                      # First run triggers onboarding wizard
# or manually:
prism advanced configure   # Set up LLM provider + API key

Run

# Interactive REPL
prism

# Autonomous mode
prism run "Find W-Rh alloys that are thermodynamically stable"

# MCP server for Claude Desktop
prism serve

See INSTALL.md for full details.

Commands

Command	Description
prism	Interactive agent REPL
prism run "goal"	Autonomous agent mode
prism run "goal" --confirm	Autonomous with tool consent
prism serve	Start as MCP server
prism search --elements Fe,Ni	Structured OPTIMADE search
prism ask "query"	Natural-language query
prism update	Check for updates
prism setup	Workflow preferences wizard
prism plugin list	List installed plugins
prism calphad status	CALPHAD installation status
prism sim status	Simulation status

REPL Commands

Command	Description
/help	Show available commands
/tools	List available tools
/skills [name]	List skills or show details
/plan <goal>	Suggest skills for a goal
/scratchpad	Show execution log
/status	Platform capabilities
/approve-all	Auto-approve all tool calls
/login	Connect MARC27 account
/save	Save current session
/load ID	Load a saved session
/export [file]	Export last results to CSV

Project Structure

app/
  agent/          # MARC27 License — Agent core, REPL, autonomous, memory, scratchpad, spinner
  skills/         # MARC27 License — 10 multi-step workflow skills
  ml/             # MARC27 License — ML pipelines and algorithm registry
  simulation/     # MARC27 License — Pyiron bridge + CALPHAD bridge
  validation/     # MARC27 License — Rule-based validation engine
  plugins/        # MARC27 License — Plugin framework
  cli.py          # MIT License — CLI entry point
  config/         # MIT License — Settings, branding, preferences
  db/             # MIT License — SQLAlchemy models and database
  data/           # MIT License — DataStore and collectors
  tools/          # MIT License — Tool definitions and registry
tests/            # MIT License — 530 tests
docs/             # MIT License — Documentation and assets

Testing

python3 -m pytest tests/ -v --ignore=tests/test_mcp_roundtrip.py --ignore=tests/test_mcp_server.py

530 tests covering agent core, tools, skills, data collectors, ML pipelines, CALPHAD integration, validation rules, plugins, and CLI commands.

Roadmap

Current (v2.1.1)

• VIBGYOR block-letter banner, braille spinner, bordered tool cards, Markdown rendering
• Approval cards with [y/n/a] (always) per-tool auto-approve
• MARC27 managed LLM access (/login)
• OPTIMADE noise-free searches (explicit provider list)
• Python 3.11+ with pyiron/CALPHAD out of the box
• Dual license (MIT core + MARC27 proprietary AI)
• Tool consent, scratchpad, plan-then-execute, feedback loops

Next (Phase G — deferred to ESA/seed funding)

• GFlowNet generative sampler
• GNN surrogate models
• Active learning loops
• Multi-agent coordination
• Playbook system

Future (Phase H — deferred)

• A-Lab robotic integration
• Federated compute
• Automated experimental validation

License

PRISM uses a dual license:

Component	License
CLI, data layer, tools, tests, docs	MIT
Agent core, skills, ML, simulation, validation, plugins	MARC27 Source-Available

See LICENSE for details. Commercial licensing: team@marc27.com

Links

• INSTALL.md — Installation guide
• SECURITY.md — Security policy
• CHANGELOG.md — Version history

---