.png){kind=icon}
.png){kind=icon}
Return
DeepDESmelt: A composition-aware mixture-of-experts model for melting point prediction of deep eutectic solvents
.png){kind=icon}
PRE
AI
G
Y
张
J
J
M
B
刘
DOI：10.1002/aic.70508.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep eutectic solvents (DESs) are pivotal green media for carbon capture, yet their melting point determination suffers from costly experiments and limited model accuracy. Herein, this study proposes DeepDESmelt, a deep learning model designed for precise and efficient predictions of DES melting points. By integrating GFP-FiLM (Gaussian Fourier projection and feature-wise linear modulation) and MoE (mixture-of-experts) modules with molecular representations from Uni-Mol2, DeepDESmelt addresses two key challenges that hinder the melting point accurate predictions: nonlinear molar ratio effects and DES chemical heterogeneity. Evaluated on a newly constructed benchmark comprising 2349 experimental DES melting points, DeepDESmelt achieves outstanding performance with R2 = 0.931 and MAE = 12.2 K on unseen DES systems. It also outperforms state-of-the-art data-driven and mechanism-driven models, yielding respective reductions in RMSE and MAE by 13.3% and 20.9%. This work provides a powerful computational tool for high-throughput screening of DES candidates with tailored liquidus windows, facilitating sustainable solvent design toward carbon neutrality.
Keywords:
deep eutectic solvent
deep learning
melting point
mixture of experts
Uni-Mol2
Journal
IF:
4
Papers:
1.1W
Citations:
2.9W
Organization
Cited Papers