Why one-time calculators fall short
Many bio-materials tools take a composition, run a model, and return a predicted property. Some provide useful starting points. Their basic limit is that they do not learn from your results. Each prediction comes from the same model and data, with no path for adding what the lab learned from its batch.
Collagen source, crosslinker purity, bioreactor geometry, and culture media can vary between labs. A model trained on published data may therefore show a consistent bias when used in one lab's workflow. Because that bias is often systematic rather than random, it can be corrected if the platform has a way to learn it.
The validation loop provides that mechanism. When a lab records measured batch properties alongside the formulation, Celleste updates the lab-specific calibration layer. Later predictions use the lab's own history as well as the broader training data.
How the loop is structured
There are three components. First, before a batch runs, the lab enters its target formulation and receives a predicted property distribution with confidence intervals. Second, after the batch, the lab records the measured properties. Third, Celleste ingests those measurements, calculates the residual between predicted and measured values, and updates that lab's calibration layer.
This is not full retraining of the underlying model after every new result. That would be expensive and could overfit recent batches. Instead, Celleste maintains a lab-specific likelihood weighting layer that adjusts the posterior distribution using accumulated residuals. The result is a correction for systematic workflow bias while the broader dataset continues to provide regularization.
Calibration converges at different rates. With replicated batches at fixed formulations, the platform can separate measurement noise from systematic bias within five to eight batches. A lab exploring a wide composition space with few replicates will take longer, because each batch samples a different region and says less about measurement noise.
What gets better over time
The first visible change is narrower confidence intervals. Initial predictions are wide because no lab-specific calibration exists. After a dozen or so batches, intervals narrow substantially for formulations resembling those already tested. The platform has learned the noise floor of that lab's measurement protocol.
Candidate ranking also improves. Absolute accuracy may remain limited, while the model becomes better at ordering formulations from best to worst for a property. In iterative optimization, that can matter more than the exact predicted value. The practical question is often which of three candidates to run first.
A third change takes longer. As calibration data accumulates, Celleste can identify which composition variables best predict results in that lab's workflow. A variable that matters strongly across the broader population may matter less where the lab's protocol already controls it. Over time, sensitivity analysis becomes sharper as calibration separates useful signal from lab-specific noise.
A pilot scenario
A team developing a collagen-hydroxyapatite composite for bone scaffolds used Celleste during a pilot. Initial Young's modulus predictions showed roughly 22% mean absolute error against measured values, with wide confidence intervals. After logging 14 batches across three compositional regimes, error on held-out batches fell to approximately 11%. The narrower intervals also let the team distinguish candidates with overlapping nominal properties.
The calibration also exposed a systematic bias: initial modulus predictions were about 15% too high for the team's formulations. The cause was traced to a difference in hydroxyapatite particle size distribution from its supplier compared with the broader training data. Once captured, the bias correction improved predictions for formulations outside the calibration set.
We do not present this as a typical outcome. Fourteen batches is a reasonable pilot size, while labs with fewer batches or sparsely sampled composition spaces will converge more slowly. The recurring pattern is that systematic bias can be corrected once the loop is closed, across different materials systems and labs.
Validation loop limits
Closed-loop calibration cannot resolve a formulation space that is genuinely novel relative to the training data. When few or no close analogs exist, even excellent lab calibration will not create narrow confidence intervals. The model still lacks the structural knowledge needed for accurate interpolation.
The loop also cannot replace disciplined measurement. Incomplete batch records or inconsistent protocols cause the update to absorb noise instead of signal. Validation checks in the data entry workflow can flag likely inconsistencies, but calibration quality remains limited by the quality of the lab's recorded data.
Finally, the loop covers formulation-to-property relationships, not process variables the lab does not record. If gelation temperature changes between batches without being logged, calibration sees noisier residuals and converges more slowly. The labs reaching tight predictions fastest are the ones recording process conditions consistently alongside composition.
