SandboxAQ's Large Quantitative Models (LQMs) combine physics-based simulation with machine learning to accelerate drug discovery and materials innovation through a proprietary three-layer infrastructure.
Now, these LQMs are accessible through leading Large Language Models (LLMs). Researchers can access frontier scientific AI though natural language prompts they already know and love, moving faster from hypothesis to discovery.
We are proud to announce our first LLM to LQM integration is live! AQCat Adsorption Spin, SandboxAQ’s spin-aware machine learning engine for heterogeneous catalysis, is now accessible via waitlist. AQCat Adsorption Spin allows researchers to lock in the most critical first step of any catalyst discovery workflow, adsorption energy calculation, so they can rapidly identify and prioritize the most promising candidates before committing costly modeling and lab resources to full-scale evaluation.
This is just the beginning! Additional models are coming soon.
Our LQMs follow a three-layer architecture: we generate physics-grounded data, train proprietary AI and physics models on that data, and chain both into automated workflows which orchestrate full design–test–make–decide loops.
We generate our own physics-grounded training data — and decide which data to generate by starting from the problem, not the method. This allows us to target the specific chemistries and operating conditions which matter most to our partners.
Many platforms are constrained by the experimental data that already exists. When public databases are sparse — as they are for novel chemistries, unexplored materials systems, and the hard targets that represent the greatest scientific and commercial opportunity — those platforms hit a wall. Because we generate our own data through physics simulation, we do not.
We run high-fidelity physics engines — Density Functional Theory (DFT), Molecular Dynamics (MD), Free Energy Perturbation, microkinetics, and more — to produce synthetic data that reflects the actual laws governing molecular behavior, not statistical patterns extracted from the literature. Where experimental data exists, we incorporate it. Where it doesn't, we can still move forward and solve the problem.
We then train our LQMs on these physics-grounded datasets— and we own the models which results. They are not licensed tools, third-party platforms, or AI wrappers over someone else's simulation engine. We build them from our own data, validate them against real-world benchmarks, and continuously improve them.
Our model portfolio includes:
We don’t stop at our own models. We are also committed to collaborating with the Open Source community and help improve public models. For example, we are proud that AQAffinity is built on– and further enhances– OpenFold3.
Our final layer orchestrates full design–test–make–decide loops with automated, agentic workflows. These workflows can encode the actual decision, under real-world constraints, and run models iteratively until a defensible answer emerges.
The agent layer powering these loops handles the cognitive load scientists shouldn't have to carry. It plans, routes, iterates, and surfaces results, while keeping scientists in control at every critical checkpoint. Our agents allow researchers to shift their focus from executing each step to making the decisions that matter.
We run these workflows across multiple domains. Drug discovery and biologics. Heterogeneous catalysis. Solid-state battery materials. Alloy design. All on a single platform, with shared infrastructure and continuously enriched data.
