Early stopping: avoid overtraining

One of the most common questions when training a model is simple: how many epochs are enough?

In practice, this often leads to overestimation. To avoid stopping too early, users tend to allocate more epochs than needed, resulting in unnecessary computational effort.

With early stopping, nvisionFlow (transient models) can now automatically stop training once sufficient convergence is reached. Instead of committing upfront, the process adapts to the data.

To give users control over this behavior, three levels are available:

• fast

• balanced

• patient

This allows you to decide how conservative you want the stopping criterion to be, while avoiding unnecessary training.

Model retraining: extend, don’t restart

But what if you stopped too early? Or what if new data becomes available?

With the new retraining feature, you no longer need to start from scratch. You can take an existing nvisionFlow model and continue training:

• by adding more epochs

• or by incorporating additional datasets

This applies to both steady-state and transient nvisionFlow models.

Retraining also complements early stopping: if the stopping was too aggressive for your use case, you can simply continue training from the existing model. The process becomes iterative, not all-or-nothing.

Direct Abaqus integration for nvisionFlow

nvision 2026.1 also improves integration into existing simulation workflows.

Users can now directly import Abaqus data for non-parametric cases. Information from .inp files is converted into shell geometries, making it immediately usable for nvisionFlow training and prediction.

This removes intermediate steps and makes it easier to bring simulation data into the AI workflow.

Export predictions for Shell Geometry workflows

For users working with the Shell Geometry interface (CSV-based), nvisionFlow now also supports exporting predictions back to CSV files.

This makes it easier to:

• post-process results

• integrate with existing tools

• or continue analysis outside nvision

A more flexible way to work with AI models

With early stopping and retraining, training becomes more adaptive and less dependent on upfront assumptions. With improved import and export capabilities, nvisionFlow fits more naturally into existing engineering workflows.

The goal is not to add complexity, but to give engineers more control:

Train when it matters, extend when needed, and integrate seamlessly with the tools you already use.

Watch more - https://youtu.be/xQCZ5iNlH0U