in what cases must fp64 be used to train the model

[ipe-zhangyz]

It seems that the default precision of DeePMD-kit is double precision, or fp64 simply. However, in the SC20 Gordon Bell Prize paper the authors showed that for water system the test errors in energy and force are the same with fp32 and fp64. I want to know if there is any system that the fp64 is a must in terms of the accuracy of a model. Maybe it depends on the system that is studied and is hard to say, but I still wander if fp32 is accurate enough for most, for example 80% or 90% of cases?

[njzjz]

We benchmark FP64 and FP32 in our new paper. It looks like they give the same accuracy.

[njzjz]

The paper only tests the precision of neural networks and its compressed form. The environment matrix is not tested.

[ipe-zhangyz]

So all the tests were conducted with environment matrix in fp64 precision, the same as the SC20 paper, right?

[njzjz]

The environment matrix is in FP64 precision. For the SC20 paper, you may contact @denghuilu.

[wanghan-iapcm]

In SC20, FP64 is used for the atomic positions and the environment matrix.

[ipe-zhangyz]

Thank you. What I really concern is why you choose to use FP64 for the atomic positions and the environment matrix specifically, and not test FP32 as you do for the neural network parts. Could you please explain it a little more? Is there any theroy that makes the atomic positions and the environment matrix must use FP64, or you just use FP64 in accordance with practice in scientific computation?