AI for Chemistry Enters the 'Configuration as Code' Era: Chemia Enables One-Click Chemical AI Model Training}

In recent years, AI has become an indispensable tool in chemistry research, widely used in property prediction, reaction optimization, and material design.

Although large models are a major trend in AI, traditional feature-engineering-based algorithms still play an irreplaceable role in many vertical fields.

Currently, beginners in scientific AI face challenges such as data processing, model tuning, and experiment reproducibility, which are time-consuming and complex.

Recently, the Shanghai AI Laboratory's Materials Science team open-sourced Chemia, a comprehensive AI model training framework designed specifically for chemical property prediction and reaction optimization.

Chemia adopts the "configuration as code" concept, encapsulating the entire process—from data preparation, feature engineering, model training, hyperparameter tuning, to property and reaction condition prediction—within a clear and intuitive YAML configuration file. With a single line of code, users can train large-scale chemical models!

The core features include:

• Powerful Algorithm Library: Over 15 classic AI algorithms, including neural networks and graph neural networks, to meet diverse research needs.
• Automated Feature Engineering: Automatically generate features like Morgan fingerprints, RDKit descriptors, and support flexible calls to pretrained models like Unimol, ChemBERTa, MolT5, etc., boosting productivity.

Intelligent Optimization Engine: Deep integration with Optuna for automatic hyperparameter tuning to find optimal model parameters.

Flexible Workflow: Supports rapid training, k-fold cross-validation, and end-to-end training-optimization processes with ease.

Github: https://github.com/flyben97/Chemia

Chemia: The "Recipe" for Chemical Research

Say Goodbye to Complex Scripts, Unlock Productivity

Chemia acts like a "recipe" for chemical research, condensing data processing, feature engineering, model training, and prediction into a clear YAML configuration. Chemists no longer need to write complex code—just modify a few lines to switch datasets, models, or features. This saves time and allows researchers to focus on scientific questions. Examples include:

Case 1: Collecting a dataset of molecular structures and HOMO-LUMO energy gaps, then training a model to predict the energy gap by simply editing the YAML file and running one command.

Case 2: Gathering a dataset of battery formulations and cycle life, then training a model to predict cycle life with minimal configuration changes, akin to following a recipe to cook a complex dish without mastering advanced cooking skills.

Extreme Flexibility for Diverse Needs

Chemia supports over 15 algorithms, including XGBoost, LightGBM, CatBoost, and neural networks. Researchers can easily compare models or optimize specific reactions, just like adjusting ingredients in a recipe to find the best flavor.

Reproducibility for Reliable Science

By saving configuration files and data, Chemia enables precise experiment replication, boosting scientific credibility and team collaboration—like sharing a recipe so others can produce the same dish exactly.

Get Started in Just Three Steps, in Five Minutes

Step 1: Prepare Your Data (CSV format)

Organize your data into a table, e.g.,

# my_data.csv

Catalyst,Reactant1,Yield

CCO,c1ccccc1,95.2

CCN,c1ccc(C)cc1,88.5

Step 2: Write Your Configuration File (config.yaml)

Specify data path, features, and target in YAML, e.g.,

experiment_name: "My_First_Reaction_Prediction"

and so on, to define your experiment parameters.

Step 3: Run!

python run_training.py --config config.yaml

All results, best models, and performance reports are generated automatically—simple and efficient!

Conclusion

Chemia was developed by PhD student Gao Ben at Shanghai AI Laboratory. For more details, check the GitHub documentation, and if you find it useful, please give it a star!

Future updates are ongoing, stay tuned.

GitHub documentation: https://github.com/flyben97/Chemia/blob/main/README.md