Supporting information for a multifidelity neural network formulation for molecular potential energy surfaces

<p>This is a supplementary information for our paper titled &quot;<em>Multifidelity neural network formulations for&nbsp;prediction of quantum chemistry potential&nbsp;energy surfaces</em>&quot;</p> <p>Supplemental information includes two data files corresponding to the&nbsp;complete sets of low and high fidelity training data used in numerical experiments. &nbsp;Format is JavaScript Object Notation (JSON).</p> <p>1. low_fidelity_training_data.json contains 74000 records</p> <p>2. high_fidelity_training_data.json contains 36988 records</p> <p>Each record consists of a numerical id (&quot;id&quot;), (x,y,z) position tuples&nbsp;(&quot;geometry&quot;) for C5H5 ordered as 5 carbon atoms followed by 5 hydrogen&nbsp;atoms, and corresponding potential energy (&quot;energy&quot;).</p> <p>Source: normal mode sampling around 2 wells, 1 transition state, and a&nbsp;set of IRCs as depicted in Figure 2.</p> <p>Usage: subsets of this data were used as needed to define different data&nbsp;amounts and different subset randomizations in Figures 5 through 8.</p>