We are not building a material. We are building the machine that finds them.
The company is Exper. The product is the loop — a closed experimental system that runs continuously and records the full conditions of everything it does. This page sets out the problem it addresses, the approach, and the path from a single cell to a platform.
The limiting factor in materials R&D is data quality.
Materials research is constrained less by a shortage of ideas than by the rate at which they can be tested and the reliability of the results. A conventional lab completes a few dozen well-controlled experiments in a quarter, and the conditions behind each result are only partly recorded.
The second limitation compounds. Each experiment is designed on the basis of earlier results, so unrecorded variation in those results is carried forward. Over successive iterations, a growing share of the outcome cannot be attributed to any known cause, and it becomes progressively harder to identify which changes produced an effect.
For the same reason, the value of an experimental dataset depends less on its size than on how completely each entry is specified. A large dataset with incomplete conditions can lead a model to learn spurious correlations; a smaller, fully specified one supports reliable inference.
Reducing uncontrolled variation is the central design objective of the loop. It is what allows each additional experiment to add information rather than noise.
The written protocol is not the executed protocol.
Reproducibility is a recognised problem across experimental science. In a survey of more than 1,500 researchers, over 70% reported having failed to reproduce another group’s experiment, and more than half had failed to reproduce their own. One significant contributor is execution: the difference between a procedure as written and as it is actually performed.
Source: Baker, M. “1,500 scientists lift the lid on reproducibility.” Nature 533, 452–454 (2016).Differences in timing, dwell, handling and sequence are rarely recorded, because they are not considered part of the method. Where such variation goes unrecorded, its effect on a result cannot be separated from the effect of the variable under test. Computer vision and activity-recognition methods can now make this variation observable from lab video, but observation alone does not remove it.
When the same system executes the procedure and writes the record, the executed and recorded protocols are identical by construction. Continuous operation also removes the variation introduced at shift changes and handoffs between operators, so runs remain comparable across days and weeks.
Experiments designed around production conditions.
Laboratory experiments are typically run under tightly controlled conditions — stable temperature, purified inputs, small batches — that do not reflect how a material is manufactured. Results obtained this way frequently fail to transfer to production, where inputs vary and process tolerances are wider.
By reproducing production-relevant variation and measuring the material’s response to it, the loop produces data on how a material is likely to behave at manufacturing scale — the information a licensee needs.
Three requirements for a reliable loop.
Experiment selection
The model must choose each next experiment from a limited, noisy history. This is an active-learning problem with a bounded cost: a poor choice consumes one cycle and still yields fully specified data.
Stable execution
Automated runs must remain repeatable over weeks of unattended operation. Gradual drift is the principal risk, because it degrades data quality without producing a visible failure.
Comprehensive capture
Every run records the full set of measurable conditions, not only those expected to matter. A variable that is not captured at the time cannot be recovered afterwards.
Together these produce an unbroken record linking each result to the conditions that produced it. That provenance is what makes a validated formulation licensable.
Predictive embedding models for physical data.
The loop generates high-dimensional data: video of each run, continuous sensor traces and measurement series. The choice of model architecture determines how much of that data is useful.
Generative models learn to reconstruct their input in full, including variation that is unpredictable and irrelevant to the outcome. Joint-embedding predictive architectures, such as Meta’s V-JEPA, make predictions in a learned representation space instead, modelling what is predictable about a scene and discarding what is not. That suits a system whose purpose is to identify sample state, detect drift in execution and attribute outcomes to conditions.
Predictive embedding objectives also learn more from each example. In a domain where every training sample is a physical experiment, and no web-scale corpus exists, the amount learned per sample is a more important constraint than model size.
From one bench cell to a platform.
Each phase ends in a defined gate. Progression depends on meeting that gate rather than on a fixed schedule.
Bench validation
A single cell with human-assisted execution on a short-cycle substrate. The objective is to confirm that the instrumentation captures the conditions that influence outcomes, and that a model-proposed sequence outperforms a fixed protocol under identical conditions.
- Full sensor capture, including ambient conditions and continuous measurement
- Baseline: a competent fixed protocol run against the same objective
- Data schema and provenance chain established before scaling
First closed loop
One cell running unattended through consecutive cycles with no intervention between them. This is the first fully closed configuration, and the first measurement of how long the system can operate before drift requires correction.
- Robotic execution from setup through measurement
- Continuous operation, with failure modes instrumented rather than patched
- Drift detection built in from the start
Parallel cells
Several identical cells running concurrently against a shared model and experiment queue. The marginal cost of an experiment falls as fixed costs are spread across more cycles, and the model improves faster because it learns from every cell at once.
- The cell becomes a standardised, replicable unit
- One queue, one model, multiple executors
- Cost per experiment becomes the primary metric
Second material class
A materially different substrate running on the same infrastructure. This phase tests whether the system generalises across material classes, which the platform model depends on.
- A new substrate productive without rebuilding the stack
- Shared model architecture across unrelated material classes
- Production-condition modelling carried across intact
Platform
The loop becomes the product: a portfolio of validated, provenance-backed formulations available to license; a structured experimental dataset offered to teams building models of the physical world; and third-party programmes running on Exper infrastructure.
- Licensable IP portfolio with full experimental provenance
- Dataset offered as a standing product
- External programmes run as tenants on the platform
Scale comes from replicating cells.
A conventional lab scales by adding staff, so costs rise roughly in proportion to output. The loop scales by replicating cells, and the marginal cost of each additional experiment falls as it does.
One enclosure, one robotic executor, one instrumentation package. Built to a standard specification and configured per material class, so additional units can be procured rather than designed.
Cells draw from one queue directed by one model, so data from each cell improves the experiments chosen for all of them.
The accumulated record forms a clean, fully instrumented stack of physical materials data — a foundation for material modelling that is not currently available at this scale.
Because each build is modular, a cell can be commissioned with its own AI-native product and engineering support: installed on site by deployed engineers, operated by a partner’s own team, or administered remotely. The configuration changes with the material; the architecture does not.
Each funding stage is tied to a milestone.
Prototype the closed loop
Complete Phases 00 and 01: a single cell operating unattended, with the instrumentation and data schema that later stages depend on. The deliverable is a working loop and an initial fully specified dataset.
Demonstrate replication and a first licensable result
Parallel cells, falling cost per experiment, and a first formulation developed far enough to license. At this stage the two revenue lines separate: data matures first, and IP follows with greater value.
Generality, then platform
A second material class on unchanged infrastructure, followed by opening the loop to external programmes. The preceding stages establish the evidence this stage depends on.
Want the detail behind any of this?
We are glad to walk through the architecture, the current phase, or what a partnership could look like.