 |
Enhancing Activation Energy Prediction under Data Constraints Using Graph Neural Networks
We developed a Graph Neural Network (GNN) framework leveraging low-cost computational data to predict activation energy with high accuracy. The delta learning method significantly reduced error (MAE: 3.85 kcal/mol) while requiring only 20–30% of high-level data. This approach offers a scalable, efficient solution for modeling activation energies in diverse chemical systems.
|
 |
Advancing vapoer pressure prediction: A machine learning approach with directed message passing neural networks
Our lab developed a machine-learning model using Directed Message Passing Neural Networks (D-MPNNs) to predict vapor pressure with high accuracy (AARD 0.617 vs. 1.36 for traditional methods). The model relies only on molecular structures and temperature data, eliminating the need for experimental inputs. This work showcases machine learning's potential to transform chemical property prediction.
|
|
Machine learning-guided strategies for reaction conditions design and optimization
This study explores the use of machine learning (ML) for optimizing chemical reactions, showcasing its potential to enhance prediction accuracy and efficiency. By combining ML with high-throughput experimentation, the research improves reaction conditions and yields, surpassing traditional methods. The integration of ML with automated platforms paves the way for self-driving laboratories and novel reaction discovery.
|
|
|
Unveiling the Role of Quantum Mechanical Descriptors in Machine Learning for Chemical Property Prediction
Our latest research explores how quantum mechanical (QM) descriptors enhance deep graph neural networks (GNNs) for predicting chemical properties like solubility, toxicity, and reactivity. We find that QM descriptors significantly improve GNN performance in data-limited scenarios, boosting accuracy and generalizability.
|
|
|
AutoTemplate—Revolutionizing Chemical Reaction Datasets for Machine Learning Applications in Organic Chemistry
We introduce AutoTemplate, an innovative data preprocessing protocol that enhances the quality of chemical reaction datasets for machine learning applications in organic chemistry. This research addresses critical challenges in data quality, leading to improved accuracy and usability for tasks like yield prediction, retrosynthesis, and reaction condition prediction. Our findings significantly advance the reliability of machine learning models in the field.
|
|
|
Advanced Deep Learning Model Enhances Chemical Synthesis Planning
This research introduces a novel two-stage deep neural network that accurately predicts optimal reaction conditions for chemical synthesis. By utilizing advanced machine learning techniques, it streamlines the process of chemical reactions, enhancing efficiency. This approach has the potential to save time and resources in laboratories globally.
|
|
|
Integrating Chemical Information into Reinforcement Learning for Molecular Geometry Optimization
This study presents a two-stage deep neural network that predicts optimal reaction conditions for chemical synthesis. By applying advanced machine learning techniques, it enhances the efficiency of chemical reactions. The approach aims to save time and resources in laboratories globally.
|
|
Machine Learning in Chemical Kinetics and Thermochemistry
This chapter explores recent advancements in applying machine learning to predict molecular thermochemical and kinetic properties. It highlights the significant impact of these techniques on understanding chemical reactions and optimizing reaction systems. The research showcases the potential of machine learning to enhance insights in the field of molecular sciences.
|
|
|
Advancing Climate Change Research with Machine Learning Models for Greenhouse Gas Prediction
This research develops machine learning models to accurately predict the radiative efficiency of greenhouse gases, which are crucial for understanding global warming. The models are trained on a comprehensive dataset generated from density functional theory and infrared spectra calculations. This approach enables efficient predictions for a wide range of halogenated and non-halogenated greenhouse gases, aiding informed decision-making in climate change mitigation.
|
|
|
New Approach to Explainable Uncertainty Quantification in Molecular Property Prediction
This research presents a novel method for quantifying uncertainties in deep learning models used for predicting molecular properties. It enhances transparency and explainability, allowing for the identification of factors contributing to uncertainty. This advancement improves the reliability of machine learning models in chemical research.
|