Synthetic Chemical Reaction

# Featured at "Prior knowledge meets Neural ODEs: a two-stage training method for improved explainability" a Tiny Paper @ ICLR 2023 https://openreview.net/pdf?id=p7sHcNt_tqo

The chemical reaction dataset is a synthetic reaction with four chemical components defined by
A+B → C+D. At the first time step there is 1g of A and 1g of B.
The goal is to predict the mass of each component over time.

There are data available to evaluate the adjusted model at 3 tasks:
-reconstruction: evaluate the performance at predicting the same time steps used for training;
-extrapolation: evaluate the performance at predicting for a longer time horizon than the one used for training;
-completion: evaluate the performance at predicting time steps in between the ones used for training.

Note that the law of the conservation of mass is being followed.

在《先验知识遇见神经ODE：一种用于提高可解释性的两阶段训练方法》一文中，作为亮点呈现的数据集为2023年ICLR会议的一篇简短论文。该化学反应数据集由四个化学成分构成，遵循A+B → C+D的化学反应方程式。在初始时刻，A和B的初始质量均为1克。研究目标在于预测各成分随时间变化的质量。数据集提供了三个评估调整后模型性能的任务数据：- 重构：评估在预测与训练所用时间步长相同的性能；- 外推：评估在预测训练所用时间步长之外的更长时间跨度的性能；- 补充：评估在预测训练所用时间步长之间的性能。值得注意的是，数据集遵循质量守恒定律。