Automated Fitting Procedures

Fitting QMD automatically is designed for developing force field parameters quickly and reliably, with minimum human operation.

Assume the default Lennard-Jones parameters are good enough to predict condense phase properties and well transferrable, the procedure fits QMD for determining valence and charge parameters only.

Four modules are implemented.

• set: adjust hyperparameter that control parameter space and fitting procedure variables.
• fit: run least-squares fit parameters against training set QM data, using both linear and nonlinear methods.
• validate: compare fit merit - the deviations between training and predicting data. Run MM computations to compare structural and energetic data between training QM data and MM simulation data.
• stability-test: parameter stability is tested by running a short NVT MD simulation at elevated "temperature".

The workflow iterates N times depending the range of hyper-parameters

1. set hyperparameter
2. fit
3. validate

In then end, pick up the best result that passes the stability-test. Output validation and test result, together with optimized force field parameters.