The paper presents a comparative analysis of two formalisms, single-generator and double-generator (GENERIC), for incorporating thermodynamic principles into neural network models for predicting physical phenomena.
The key highlights and insights are:
Both formalisms show high accuracy in results that are coherent with the laws of thermodynamics, even when the single-generator formalism does not explicitly impose the thermodynamic constraints.
The single-generator formalism generally yields better results with lower computational cost, but its stability depends more on the proper tuning of hyperparameters like learning rate and network capacity.
The GENERIC formalism, with its separation of reversible and irreversible dynamics, is more representative of the underlying physics, but the added degeneracy conditions introduce an extra hyperparameter that needs to be balanced with the data loss.
The GENERIC formalism exhibits higher robustness to changes in the database and hyperparameters compared to the single-generator approach.
The performance of the formalisms also depends on the characteristics of the physical problem, with GENERIC showing advantages for weakly dissipative dynamics, while the single-generator approach is better suited for highly dissipative systems.
The paper provides insights on the trade-offs between the expressiveness and learnability of the network when incorporating different levels of physical constraints.
To Another Language
from source content
arxiv.org
Глибші Запити