Neural network ensemble for computing cross sections of rotational transitions in H₂O + H₂O collisions

Area of Science:

• Astrophysics
• Computational Chemistry
• Machine Learning

Background:

• Water (H2O) is abundant in astrophysical environments.
• Modeling H2O + H2O collisions is crucial but computationally intensive.
• Existing quantum mechanical methods are intractable for complex molecular systems.

Purpose of the Study:

• Develop an efficient machine learning tool for predicting rotational transition cross sections in H2O + H2O collisions.
• Construct a database of rate coefficients for astrophysical modeling.
• Overcome computational limitations of traditional quantum mechanical methods.

Main Methods:

• Utilized a machine learning tool with an ensemble of neural networks (NNs).
• Employed data computed with mixed quantum-classical theory (MQCT).
• Trained NNs on approximately 10% of computed data to predict cross sections.

Main Results:

• NNs accurately interpolated 12 quantum numbers for rotational transitions.
• Achieved an average relative root mean squared error of 0.409 in cross-section prediction.
• Thermally averaged cross sections showed excellent agreement with MQCT calculations (∼13.5% deviation).

Conclusions:

• The developed ML methodology is robust and efficient for predicting molecular collision rates.
• This approach can be applied to other complex molecular systems in astrophysics.
• Enables accurate modeling of water-rich astrophysical environments.