When engineers use neural networks to predict airflow around cars, the prediction often changes if the car is rotated or measured in different units. This is a problem because physics says the result should be the same regardless of how you describe the problem.
ISLA is a new neural network that never uses raw coordinates. Instead, it measures distances and angles relative to learned anchor points, using a ruler that automatically scales with the object. Every calculation respects the true area of each surface cell, not just the number of points.
On a test with 2,816 different flows, ISLA's errors were at least 3.1 times lower than any previous method. It also maintained near-perfect accuracy when the car was rotated or the mesh was changed, making it reliable for real-world engineering use.