Robustness and accuracy improvement of data processing with 2D neural networks for transient absorption dynamics

Area of Science:

• Physical Chemistry
• Spectroscopy
• Computational Chemistry

Background:

• Conventional methods for analyzing transient absorption spectra face limitations in fitting ability and robustness, especially with noisy data.
• Accurate spectral analysis requires tools that can predict intrinsic spectral properties without arbitrary assumptions, even under challenging conditions.

Purpose of the Study:

• To develop a new, accurate, and robust algorithm for analyzing transient absorption spectra.
• To overcome the limitations of conventional methods in handling noisy backgrounds and complex spectral data.

Main Methods:

• Integration of Lasso regression and neural network models for spectral fitting.
• Development of a network capable of automatically determining exponential forms for spectral units.
• Utilizing neural networks for predicting lifetimes and amplitude ratios on a per-wave unit basis.

Main Results:

• Achieved up to 97% accuracy in spectral fitting by automatically determining exponential forms.
• Enabled accurate prediction of lifetime and amplitude ratios, overcoming computational limitations of global fitting.
• Demonstrated significant improvement in fitting accuracy under weak signals, with mean square error (MSE) decreasing over 100 times compared to conventional methods.

Conclusions:

• The developed algorithm offers superior accuracy and robustness for transient absorption spectral analysis.
• The method effectively handles noisy data and weak signals, providing reliable predictions.
• This approach is readily applicable to time-resolved transient spectra analysis, enhancing scientific discovery.