Table 1.
Comparison of the standard procedure with our iterative training approach
| Standard procedure (ref. 35) | Iterative training (this paper) | |||
|---|---|---|---|---|
| Tf | # TLS | # NEBs | # TLS | # NEBs* |
| 0.062 (Ar) | 61 | 14,202 | 156 | 2500 |
| 0.062 (NiP) | 28 | 14,202 | 59 | 2000 |
| 0.07 (Ar) | 291 | 21,109 | 1057 | 5000 |
| 0.07 (NiP) | 46 | 21,109 | 152 | 4000 |
| 0.092 (Ar) | 245 | 30,920 | 776 | 4000 |
| 0.092 (NiP) | 28 | 30,920 | 123 | 6000 |
Analysis of data collected in ref. 35. We report results for different glass stabilities, decreasing from top to bottom, using Argon units (Ar)30 and NiP metallic glass parameters (NiP)27. The standard procedure finds less than half of the TLS found with the ML and is computationally much more expensive.
*This number does not include the data that we use to pre-train the model.