Studio · ML-PREDICT & ML-REVERSE

Predict process outcomes, or invert to recommend settings.

ML-PREDICT learns the forward map from process parameters to outcomes. ML-REVERSE inverts it. Given a target, it returns the settings most likely to hit it. Both are physics-informed surrogates of the CO-SIMU library, trained on synthetic data the CO-SIMU models generate and calibrated against your own measurements.

Request an ML scoping call →See CO-SIMU

What ML-PREDICT and ML-REVERSE are

Surrogates trained on the physics, calibrated on your line.

ML-PREDICT and ML-REVERSE are narrow, high-accuracy machine-learning models that approximate the CO-SIMU mesoscopic simulators. They take seconds to evaluate where the underlying physics takes hours. Because they are trained primarily on the physics, they stay physically grounded where measurements are sparse.

The same architecture supports both directions. Trained one way, the model predicts outcomes from process parameters. Trained the other way, it inverts the problem and recommends settings to hit a target, always within an explicit perimeter of validity.

The ML-PREDICT & ML-REVERSE capability stack

Four capabilities, one consistent architecture.

01 / 04

Forward prediction (ML-PREDICT)

From process parameters (slurry composition, drying profile, calendering pressure, infiltration conditions) predicts outcomes such as porosity, conductivity, tortuosity, saturation, yield and capacity loss. Inference completes in seconds.

02 / 04

Inverse recommendation (ML-REVERSE)

Given a target performance envelope, returns the process settings most likely to hit it. The inverse map is trained as its own model, which avoids random search over the forward model. Each recommendation is reproducible and carries its audit trail.

03 / 04

Hybrid physics + ML

Trained on synthetic data generated by the CO-SIMU mesoscopic simulators, then calibrated against your real measurements. The physics constrains the surrogate where your data is thin. Your measurements pin it down where the data is dense.

04 / 04

Confidence & safe rails

Every prediction carries an explicit confidence score. A self-diagnostic module flags inputs that fall outside the trained perimeter.

ML-PREDICT & ML-REVERSE at a glance

hrs → s

from CO-SIMU runtime to surrogate inference

2 ways

trained for forward and inverse direction

Q4 2026

first release available

Where they get used

From rheology to saturation. Concrete forward maps.

On mixing and coating, the forward model learns slurry parameters → viscosity and rheology. On drying, it maps slurry composition and furnace parameters → porosity, conductivity and tortuosity, with classification heads for defects such as cracks. On calendering, pressure and roll speed → porosity, conductivity and tortuosity. On infiltration, viscosity, density, contact angle, separator microstructure, pressure and temperature → saturation percentage.

Each of these maps can be inverted. Given a target porosity, saturation or capacity, ML-REVERSE returns the parameter window along with a confidence score, so engineers move from outcome objectives to actionable setpoints in a single step. Bayesian optimization layered on top picks the most informative experiments when you do need to extend the perimeter.

Common questions

Data, deliverables and timing.

No. The bulk of the training data is generated synthetically by the CO-SIMU mesoscopic simulators. Your role is to provide a small set of real measurements (typically five to ten samples covering the perimeter you care about) to calibrate the surrogate against your reality. If the perimeter is wider, a short experimental design extends the training set.

A regression on industrial data alone over-fits sparse regions and silently extrapolates into territory it has never seen. ML-PREDICT is built on physics-based synthetic data from the CO-SIMU library, so it stays meaningful where measurements are thin. It also carries an explicit confidence score that flags out-of-perimeter inputs.

A library or binary containing the ML weights and inference code, or a hosted deployment on your cloud or on-premise servers. Choice of delivery follows your data governance and integration constraints. Both modes are fully supported.

First release is scheduled for Q4 2026. Customer-specific scoping and CO-SIMU calibration work can start now, so the model is calibrated and trained on day one of availability.

ML-PREDICT & ML-REVERSE

Move from hours of simulation to seconds of inference.

First release ships Q4 2026; scoping starts now. We plan the CO-SIMU training runs, the calibration data we need from you, and the ML-PREDICT or ML-REVERSE deployment.

Request an ML scoping call →See CO-SIMU