fig4

Figure 4. Percentage difference between the five-fold mean RMSE obtained from re-execution and that recorded in the original MatBench submission, calculated using Eq. (1). Positive values indicate that the re-executed RMSE was higher than the reported value. In contrast, negative values indicate that it was lower. Values of 0 indicate that the reported and re-executed five-fold mean RMSE values were identical within the numerical precision of the stored results. Blue and red bars represent composition-based and structure-based tasks, respectively. RMSE: Root mean square error; RFLR: Random Forest by Lorenz Romaner; coGN: Connectivity optimized Graph Network; coNGN: Connectivity optimized Nested Graph Network.