How to add real sensor data to PINN

[peg33608]

I want to train a thermal flow model.
Now I have the geometric model of the machine and a lot of temperature data from various sensors, but I lack initial velocity data.
How can I incorporate these data into the training process using the Physics-Informed Neural Network method?

Thank you.

[daniel.fleischer1]

This might help:

How to create a digital twin with only data-driven?

Hi @matiyanez Yes, Modulus does allow for physic-driven, hybrid, and data-driven training. An example of a pure data-driven problem is the Darcy problem which has image like data. For point wise data you can use the PointwiseConstraint.from_numpy() where you can feed in a set of numpy dictionaries to train from. This is used in a couple spots of our examples such as the three fin heat sink which can combined with other physics-based training constraints.

[peg33608]

Thank you for your reply.
I have read those examples, but I am still confused.
I have many different geometry and corresponding temperature data, but I lack any initial velocity data.
This missing velocity data makes it challenging to calculate the PDE that requires velocity, as seen in the Darcy example using the PINO method.
On the other hand, if I use the three-fin example, which is only designed for a single geometry, can I only use one pair of data within pointwise constraints for training the model?