SandboxAQ’s Physics‑Based AI Now Runs Scientific Simulations Through Claude

Turning Claude Into a Scientific Simulation Interface

The integration links Claude to SandboxAQ’s proprietary simulation models through MCP, an open protocol designed to connect AI assistants with external tools and data systems.

In practice, this means a researcher can describe a scientific task—such as screening catalysts or evaluating molecular behavior—in natural language. Claude interprets the request and sends it through the MCP connection to SandboxAQ’s models, which run the underlying calculations and return results.

This workflow replaces what traditionally required custom scripts, complex APIs, and access to specialized computing environments. According to SandboxAQ’s documentation, scientists can run high‑accuracy molecular simulations conversationally without managing infrastructure or writing integration code.

What Makes Large Quantitative Models Different

SandboxAQ’s LQMs differ from standard large language models. Instead of learning patterns from text, they are physics‑grounded models designed to simulate real‑world systems, including chemical reactions, molecular dynamics, and other quantitative processes.

The company describes LQMs as combining physics‑based simulation methods with machine learning to accelerate discovery in areas like drug development and materials science.

By pairing these quantitative models with a natural‑language interface, the system allows researchers to interact with scientific simulations the same way they might query an AI assistant.

First Use Case: Catalyst Discovery With AQCat Adsorption Spin

The first tool available through the Claude integration is AQCat Adsorption Spin, a model designed for heterogeneous catalyst discovery.

Through natural‑language prompts, materials scientists can:

• Predict adsorption energies for molecules on catalyst surfaces
• Identify binding‑energy crossover points
• Screen catalyst compositions for promising reactions

The model is built on SandboxAQ’s spin‑aware machine‑learning engine for catalytic systems and can provide insights similar to density functional theory (DFT) simulations but with far less manual setup.

Catalyst discovery is a major application area because catalysts underpin large portions of industrial chemistry and energy production. Data used to train models such as AQCat25 include millions of quantum‑chemistry calculations across tens of thousands of catalyst systems.

Next Targets: Drug Discovery Models

SandboxAQ says additional models will follow the same conversational access pattern.

Two upcoming LQMs highlighted for biopharma workflows are:

• AQPotency, focused on predicting the potency of candidate drug molecules
• AQCell, designed to model cellular interactions and biological effects

The goal is to bring computational drug‑discovery tools—traditionally accessible mainly to specialized computational chemistry teams—into a broader research workflow powered by natural language interfaces.

Why SandboxAQ Calls This a Step Toward “Quantitative AI” at Scale

The company frames the Claude integration as more than a single product feature. Instead, it represents a broader strategy: distributing physics‑based AI models through widely used language‑model interfaces.

Historically, running advanced scientific models required both domain expertise and programming skills. By using LLMs as the interface layer, SandboxAQ aims to reduce that friction so researchers can focus on scientific questions rather than software integration.

If widely adopted, this approach could extend quantitative simulation tools across industries including:

• Biopharma and drug discovery
• Energy and catalyst development
• Industrial chemistry
• Advanced materials research

The underlying idea is that researchers can move more quickly from hypothesis to simulation to insight when the interface barrier drops from code to plain language.

What This Means for AI‑Driven Scientific Research

The integration highlights a growing pattern in scientific AI: pairing language models as orchestration layers with specialized models that perform real scientific computation.

Instead of asking an LLM to reason about chemistry or materials purely from text, Claude can now route requests to physics‑aware models built specifically for those domains. The result is a hybrid workflow where conversational AI coordinates simulations that operate on the actual mathematics and physics of molecules and materials.

If successful, systems like this could make advanced computational science accessible to far more researchers—without requiring them to become experts in programming, simulation pipelines, or high‑performance computing.