Thank you very much for your answer! So, I believe that learning molecular force fields is simply a process of adding and removing noise, and it does not involve the potential energy attribute data of molecules in the dataset, which should be correct. I have another question. From the paper you provided, I noticed that the denoising loss is used as an auxiliary loss. Therefore, to obtain a molecular representation that includes dynamic characteristics, a main task is still required, which is to predict a value y. The denoising loss is used as an auxiliary optimization for the predicted value y. Does this y value only represent force or energy? If applied to other types of datasets that do not contain energy-related data, what could this y value be? Or could it be another value, such as predicting the original atomic coordinates? Is that reasonable?
Thank you very much for your answer! So, I believe that learning molecular force fields is simply a process of adding and removing noise, and it does not involve the potential energy attribute data of molecules in the dataset, which should be correct. I have another question. From the paper you provided, I noticed that the denoising loss is used as an auxiliary loss. Therefore, to obtain a molecular representation that includes dynamic characteristics, a main task is still required, which is to predict a value y. The denoising loss is used as an auxiliary optimization for the predicted value y. Does this y value only represent force or energy? If applied to other types of datasets that do not contain energy-related data, what could this y value be? Or could it be another value, such as predicting the original atomic coordinates? Is that reasonable?