Behind generative formulation design
Celleste uses inverse design principles from computational materials science for cultivated bio-material formulation. This page describes the methodology.
Inverse design: from target properties to compositions
Traditional formulation proceeds forward: select a composition, then measure its resulting properties. Inverse design reverses that sequence. You specify a target property profile, and the algorithm searches composition space for formulations likely to deliver it. Celleste runs a constrained generative search across composition variables, guided by a property prediction model trained on literature priors and your lab's batch data.
Property prediction
Celleste's core model uses Gaussian process regression and domain-informed feature encoding for bio-material compositions. For every candidate composition and property, it returns a predicted value with a calibrated uncertainty estimate.
Generative search
A constrained Bayesian optimization loop proposes candidate formulations, balancing exploration of novel compositions against exploitation of high-confidence property-space regions.
Closed-loop learning
When you log a real batch result, model uncertainty narrows in that part of composition space. As data accumulates, the system becomes more specific to your material class and lab conditions.
How we assess prediction reliability
The prediction model is calibrated against held-out batch data from our pilot formulations. Calibration means its confidence intervals match observed accuracy: when Celleste reports an 80% confidence interval, the true value falls inside it 80% of the time in our validation set. We openly report calibration metrics to pilot users, allowing you to assess whether the confidence levels represent real uncertainty in your material system.
Validation dataset: 3 material systems and 47 labeled batch records from our 2025 internal pilot. Studio and Research plan users can request calibration scores and methodology.
Source: 2025 internal pilot validation. Bootstrapped-stage figures, not large-scale benchmarks.
At the intersection of computational materials science and bio-materials
Our work combines two research foundations: inverse design methods developed for inorganic materials discovery over the past decade, and compositional structure-property relationships examined in bio-polymer and cultured material research. We are not taking a general machine learning framework and applying it to a new domain. Feature representations and uncertainty quantification were built for the composition variables and failure modes specific to cultivated bio-materials.
Meet our Team