We closed our angel round in December 2025, raising $1M. It has been about six weeks since then, and the honest answer to "what changed" is: not as much as you might expect on the product side, because we were already deep in the work, and more than you might expect on the team and focus side, because having a funding event forces you to get explicit about what you are actually building and for whom.
This post is a brief, factual update on where we are across the three areas that matter most for us right now: the MLIP model, the synthesis API integration, and the early-access program. No projections, no roadmap speculation. What we have built, what is working, and what is not yet working.
MLIP Model Iteration
The core of what AstraIQ does computationally is screen inorganic candidate libraries using machine-learned interatomic potentials. We have been iterating on model architecture and training data since the company was founded in 2023, and the past three months have been focused on two specific improvements: improving coverage of cathode-relevant composition spaces and reducing the false-positive rate on thermodynamically plausible but synthesis-inaccessible structures.
Coverage expansion
We pulled additional DFT reference calculations from the Materials Project and OQMD to extend our training set coverage into Mn-rich NMC compositions and several Li-excess layered oxide families that were underrepresented in our earlier models. We held out a test set drawn from those families before training and used it to verify that per-structure energy MAE on those composition classes improved to within our target threshold. It did. The model we have now is measurably more accurate on Mn-rich cathode families than what we were running three months ago.
We have not expanded into sulfide electrolytes yet. That requires a meaningfully different training set, and we have decided to treat it as a separate model rather than trying to cover it with the oxide model. We expect to start that work in the next several months.
False-positive filtering
One of the persistent complaints about MLIP-based cathode screening is that the screens produce candidates that look good on paper (plausible formation energy, promising compositional intuition) but fail to form as phase-pure materials under any synthesis conditions the team has access to. Some of this is genuinely a model accuracy issue. But we have found that a significant fraction of these failures come from candidates that are thermodynamically stable but kinetically inaccessible under solid-state synthesis conditions: they require either very high temperatures that cause decomposition of other phases in the system, or very slow cooling rates that are impractical at scale, or precursor combinations that react in undesired ways before reaching the target phase formation temperature.
We have been developing a synthesis feasibility scoring model that encodes some of these kinetic constraints as heuristics derived from our synthesis database and from published solid-state chemistry literature. The model is still early-stage, meaning we trust it as a relative ranking tool within a composition family rather than as an absolute predictor. We apply it as a soft filter in the synthesis dispatch queue, giving lower synthesis priority to candidates with low feasibility scores rather than excluding them entirely. Our internal tests suggest it reduces the fraction of synthesis slots used on practically inaccessible candidates, but we do not yet have enough synthesis throughput data to publish a confident number.
Synthesis API Integration Work
The value proposition of AstraIQ depends on being able to close the loop: screen computationally, send top candidates to automated synthesis, get results back, and repeat. The synthesis side of that loop requires integrating with laboratory automation hardware, and that integration is harder than the computational side in ways that are somewhat specific to the solid-state chemistry context.
Where we are with hardware integration
We have a working API integration with one solid-state synthesis automation platform that handles precursor dispensing, furnace scheduling, and basic XRD characterization dispatch. We are not naming the platform or the integration partner because the integration is in active development and we do not want the description to be out of date two weeks after this is published.
What the integration currently does: AstraIQ can dispatch a structured synthesis job (structure specification, synthesis route, temperature profile, atmosphere) to the synthesis queue, receive a synthesis status update when the run completes, and ingest an XRD pattern result for phase identification. Phase identification currently uses a reference library matching approach with human review for ambiguous patterns. Fully automated phase identification is a goal but not our current state.
What is still manual
Precursor inventory management is still largely manual on the synthesis side. The automation platform handles dispensing from available stocks, but deciding when to reorder precursors, handling precursor purity variability across batches, and substituting when a specific precursor is out of stock are all handled by the lab team. We have not tried to automate these yet because the failure modes of getting them wrong are consequential (wrong precursor chemistry, inconsistent results across synthesis runs) and the gain from automation is smaller than the gain from better screening and dispatch.
Characterization beyond XRD is also still largely manual. SEM imaging, EDS mapping, and electrochemical characterization for promising candidates are done by the synthesis team on a case-by-case basis after XRD confirms phase identity. We are not planning to automate these in the near term.
Early-Access Program
We started conversations with materials R&D teams in late 2025 to understand how the platform fits into real workflows. We are working with a small number of pilot groups who are trying the screening pipeline on their specific material classes.
The feedback from those conversations has been consistent on two points. First, the computational screening is useful, particularly for teams that do not have dedicated computational materials science headcount and are currently relying on intuition and literature search to pre-filter candidates before synthesis. Second, the synthesis integration is the feature they care about most, and the current state of the integration is not yet complete enough for them to depend on it as a primary workflow. That is honest feedback and it is directionally correct: the synthesis integration is where we are investing most of our engineering effort.
We are not publishing early-access participant names or providing usage statistics because we are at an early stage and the numbers would not be meaningful. If you are at a materials R&D group and are interested in the early-access program, the contact page is the right starting point.
What We Are Not
A few things worth stating clearly, because we have been in enough conversations to know they need to be said. We are a three-person team. We have one angel investor. We are pre-revenue. We are not running at scale for a large customer base; we are building and refining the core product loop with a small number of early-access partners.
We are also not a general-purpose materials informatics platform. We are specifically focused on the inorganic screening to automated synthesis loop for crystalline solid-state materials. We do not support organic synthesis. We do not support molecular property prediction. We do not support polymer informatics. If those are your needs, there are better-suited tools for them.
What we are building is a specific capability: take a candidate library of inorganic crystalline structures, screen them computationally with calibrated accuracy, send the survivors to automated synthesis without a manual handoff, and route the synthesis results back into the next screening iteration. If that matches your R&D bottleneck, we want to talk. If it does not, we would rather say so clearly than overpromise scope. More technical detail on what the platform does and how the screening science works is on those pages if you want to go deeper.