--- url: https://invertbio.com/product/models path: /product/models kind: marketing --- # Train, version, and deploy models on your process data. Every process model your team needs for optimization, simulation, and monitoring. From soft sensors, to forecasts, control charts, mechanistic and hybrid models, and even classifiers. Process models are available for you to leverage on your data to its fullest, versioned like code, and deployable for consistent usage. [Book a demo](/demo) Train ## Unlock the potential of your process data. Models can incorporate all your process data, from online time-series, offline samples, to events and metadata. Pick your desired use case, your desired target, and develop a reviewable and auditable model training record. [See the data layer underneath](/product/foundation) Invert — Train model · Titer forecast Training set Run family CHO-Fed-Batch · Ambr250 Runs matched 48 runs · 4 campaigns Features DO %pHCapacitanceOURFeed rate+7 Target Harvest titer (g/L) Model form Hybrid · kinetics + learned residual Holdout: 12 runs · campaign 4 Fit · predicted vs. observed FIT COMPLETE · 41 s 0.94 R² holdout 0.18 g/L RMSE 48 runs trained on Version ## Versioned like code, traceable like a batch record. Each model tracks the runs and data used for training, the features or parameters, and the scoring metrics in a Model Development Record. Promote a version to production once reviewed and verified. A reviewer can always see what changed between v3 and v4, and why. Model registry Models›Titer forecast · CHO fed-batch ComparePromote Version Status Trained on Holdout R² Owner v4 Staging 48 runs · through C4 0.94 +0.06 A. Okafor v3 Production 36 runs · through C3 0.88 A. Okafor v2 Locked · C2 24 runs · through C2 0.81 J. Lindqvist v1 Archived 18 runs · bench only 0.72 J. Lindqvist v3 → v4 +12 runs from campaign 4 +Capacitance added as a feature −Seed-train density dropped (no coverage at 2,000 L) Deploy ## Score the batch that is running now. Apply a deployed model to any new run as the data arrives. No notebook to re-run, no analyst in the loop. Soft sensors fill the gap between offline samples. Forecasts update as the batch progresses. Monitoring models raise an alert and can jumpstart troubleshooting as issues arise. [See how Assist uses your models](/product/intelligence) Invert — Deployed models · B-114 LIVE B-114 · h62 elapsed CHO-Fed-Batch · 3 models scoring LIVE Titer forecastv3 3.05 g/L at harvest 95% PI 2.81–3.29 · goal 3.4 g/L VCD soft sensorv7 31.2 ×10⁶/mL Updated 4 s ago · last offline h56 Multivariate SPCv2 T² above 99% limit Top contributors: feed rate, OUR, DO h0ALERT · h62h336 ## Every model comes with its own development record. The Model Development Record holds the code that produced the fit and the decisions taken along the way. Why a feature was added. Why a holdout changed. Who reviewed it, and what they asked for. A reviewer reads the record itself, not a story reconstructed from memory months later. Model Development Record · Titer forecast v4Staging Fit code · v4 ``` # mdr/titer-forecast/v4.py from invert import runs, models train = runs.query( family="CHO-Fed-Batch", campaigns=["C1", "C2", "C3", "C4"], ).exclude(status="aborted") model = models.Hybrid( kinetics="monod", residual=models.MLP(hidden=[32, 16]), features=["DO", "pH", "capacitance", "OUR", "feed_rate"], target="harvest_titer", ) fit = model.fit(train, holdout="C4", seed=7) fit.report() # R² 0.94 · RMSE 0.18 g/L ``` Signed commit · reproducible from the pinned run set Decision log 1. Mar 04A. Okafor Added capacitance as a feature. Offline VCD alone under-predicted the late exponential phase on every C3 run. 2. Mar 06A. Okafor Dropped seed-train density. There is no coverage at 2,000 L, so the term would not transfer to production scale. 3. Mar 11J. LindqvistReviewer Holdout changed to campaign 4 only. Random splits leaked within-run correlation and inflated R² by roughly 0.05. 4. Mar 12J. LindqvistReviewer Approved for staging. Re-review before production once six more runs at 2,000 L are available. 4 decisions logged2 reviewersPinned to 48 runsExportable for review ## Every process model your team needs. Invert trains, versions, and deploys the full range of models: soft sensors, forecasts, control charts, equivalence tests, mechanistic and hybrid models, and classifiers. Here are eight examples of process models you might find useful. ### Batch monitoring and alerting with multivariate SPC Is the batch running right now drifting from normal? T² = tᵀS⁻¹t · SPE = ‖e‖² A reference PCA trained on an in-family batch set, scoring the live batch. T² crosses the limit at hour 62, and contributions rank the parameters behind it. ### Real-time cell density from a PAT soft sensor What is the cell density right now, between offline samples? VCD = f(capacitance, OUR, feed) PREDICTED VCDOFFLINE SAMPLES Regression on capacitance, OUR and feed rate, fitted against offline counts, shown against the offline measurements. ### Scale-up from a mechanistic growth model If the cells are growing like this at 5 L, what happens at 2,000 L? dX/dt = μ(S)·X − k\_d·X Kinetic parameters fitted to bench fermentations, then solved at production volume. The solid curve is the fit. The dashed curve is the prediction at 2,000 L. ### CAR-T site comparability by statistical equivalence Does the second site make the same product? −δ ≤ 90% CI(Δ) ≤ +δ 90% confidence intervals on the site A to site B difference. Equivalence holds where the interval sits inside ±10%. ### Mid-run titer forecasting with a hybrid model Where will this batch land, and is that still on goal? dX/dt = μ(S)·X − k\_d·X + NN(z) Kinetic equations with a learned residual term, solved forward from the current timepoint. The shaded band is the 95% prediction interval, here sitting below the goal line. ### AAV full-capsid risk classification with machine learning Which transfections will come in low on full capsids? P(low % full) = g(x h0–24) A classifier trained on past outcomes, scoring the first 24 hours after transfection. The dashed line is the decision boundary, and feature attributions explain each call. ### AAV affinity capture stability with I-MR control charts Is the process stable, and which batch broke the pattern? UCL = x̄ ± 2.66·MR̄ Individuals plotted against limits from a locked baseline. A point past the UCL flags the shift, and a single moving-range spike separates a step change from a drift. ### Elution optimization with a chromatography model Where should we cut the pool to hold purity and keep yield? ∂c/∂t = −u·∂c/∂x + D·∂²c/∂x² A transport model of the column, solved across candidate cut points. The shaded window is the pool that holds 0.3% HCP at 91% step yield. ## From runs to a model in production. 01 ### Train your model Your data is now connected with data from process instruments, historians, and offline samples into run-centric canonical metrics. Select your relevant runs and desired target, and build your model. 02 ### Version, review, promote Each fit is a version, pinned to its training runs, parameters, and holdout metrics. Compare two versions side by side, promote one to production, and lock a version for a campaign so the numbers in a report stay reproducible months later. 03 ### Deploy and use your model with live runs Promoted models can be applied to live runs or to historical data for simulation, optimization, experimental design, and process monitoring. Continue exploring ## Models are only as good as the data beneath them. [Foundation ### One data layer. One vocabulary. One training set. Instruments, historians, LIMS, and ELN resolved into canonical, run-centric metrics with full lineage. That is why a model trained today still means the same thing next campaign. Explore Foundation](/product/foundation) [Intelligence ### Ask your models questions in plain language. Assist reads your deployed models alongside your runs, so “why is B-114 tracking low?” comes back with the forecast, the contributing parameters, and the auditable Python behind it. Explore Intelligence](/product/intelligence) ## Put your process models to work. Most process models live in a notebook on one scientist's laptop and never score a live batch. Avoid missing out on the learnings from data you've already generated. Turn your data into insights through model training, versioning, and deployment, starting with the examples above. [Book a demo](/demo)[Start with the data layer](/product/foundation)