Request for Startups: AI for Scientific Discovery

We want founders using AI to solve problems in biology, chemistry, and physics. We are interested in models whose outputs become new molecules, better materials, and working physical systems.

The opportunity is to make scientific progress less dependent on expensive trial and error. A model might propose a promising design, predict how it will behave, or choose the experiment that would teach us the most. The resulting measurements should help improve its next decision.

For this RFS, we are looking for companies where that scientific capability is the core product. These are six directions we would like more founders to pursue.

AI designed drug molecules

We want founders building models that propose molecules for a specific disease target and help decide which ones to make and test. The opportunity includes finding new starting points for drug discovery and improving the properties of an existing compound. Research such as TamGen explores molecular generation guided by the target protein.

A useful starting point is one target with a reliable laboratory test. The model could search for compounds that act strongly on that target while limiting activity against related proteins. It should also account for whether the proposed molecules can be made.

The first evidence should come from newly tested compounds, with laboratory results compared against predictions. A favourable binding prediction alone does not establish that a molecule will become a medicine. A company could build around the resulting compounds and the ability to discover better ones.

Enzymes designed for industrial jobs

Enzymes are proteins that accelerate chemical reactions. We want founders using generative models to design them for a specific job, such as processing a difficult raw material or operating under conditions where an existing enzyme performs poorly.

This is already an experimental research direction: RFdiffusion2 generated enzyme designs that researchers subsequently tested for catalytic activity. The startup opportunity is to turn a capability like this into useful performance in a defined industrial process.

Start with one reaction and a measurable limitation, such as insufficient activity or a short operating lifetime. Generate candidates, test them, and use those results to guide another round. The eventual product could be an enzyme formulation or a production process built around it. Its value would depend on performance under the customer's actual conditions.

AI that discovers better battery chemistry or better materials

We want models that help discover battery formulations suited to a particular use, such as repeated fast charging or operation at higher temperatures. One starting point is the electrolyte, the material through which ions move inside a battery.

The model would learn from experimental results, propose new formulations, and choose which combinations to test next. Research combining robotic experiments with machine learning has demonstrated this approach for electrolyte discovery.

An early company could focus on one battery chemistry and one performance goal with a testing partner. The challenge is to improve that goal while accounting for competing requirements such as lifetime, stability, and cost. Success could produce a valuable formulation and a growing body of experimental data that helps the model make better choices.

Physics models that accelerate hardware design

We want founders building AI models that predict physical behaviour accurately enough to help engineers explore more designs. A focused application could be heat flow through a cooling system, fluid flow through a channel, or stress in a particular class of components.

Neural operators are one approach: they learn to approximate solutions to families of physics equations. Faster predictions could make it practical to compare many candidate designs before committing to detailed simulations and prototypes.

A strong starting point is one engineering problem with a clear range of operating conditions. The model should demonstrate useful accuracy on unfamiliar designs within that range, identify when its predictions are unreliable, and support a design decision that can be checked against a trusted solver or physical measurement. The business would rest on a scientific modelling capability with a measurable advantage in that application.

AI that chooses and runs scientific experiments

We want systems that take a scientific objective, choose an informative experiment, interpret the result, and decide what to try next. A scientist would define the objective and operating limits; the system would handle an increasing share of the experimental decisions.

Work such as Coscientist has explored AI-assisted experimental design and execution in chemistry. We are interested in extending this kind of capability within a well-defined field.

A first system could optimise one class of chemical reactions using existing laboratory equipment. It should keep track of what was actually performed, distinguish a failed measurement from an unpromising result, and update its choices accordingly. Compare its progress with a credible baseline under the same experiment budget. The value is reaching a useful scientific result with fewer experiments or less elapsed time.

AI that designs optical components

We want models that start with a desired optical behaviour and generate a physical structure that could produce it. This is inverse design: specifying what a component should do, then searching for its shape or structure.

Possible applications include optical filters, light-guiding components, and surfaces with particular absorption properties. Experimental work on AI-designed photonic surfaces shows how model-generated designs can be fabricated and measured.

An early company could focus on one component family and work with a fabrication partner. The model would need to account for manufacturing limits and variations, then learn from differences between predicted and measured performance. A useful outcome could be a component with improved efficiency, a smaller footprint, or a response that is difficult to obtain with existing designs.

What we want from founders

Tell us which scientific problem you understand, what your model would contribute, and how you would test its output. We want to see the connection between the AI and a result that matters to someone working in that field.

A small team can begin with an existing model, a carefully chosen dataset, and access to the right experimental partner. Its contribution should become clear through the quality of its predictions, the experiments it selects, or the things it designs.

Variance is a free, no-equity, 30-day residency in Bengaluru. Bring a specific question you want to answer and a plan for making progress on it during the month.

If you are building in one of these directions, apply to Variance.