Research ArticleRESEARCH METHODS

Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments

See allHide authors and affiliations

Science Advances  13 Apr 2018:
Vol. 4, no. 4, eaaq1566
DOI: 10.1126/sciadv.aaq1566
  • Fig. 1 Schematic depiction of a paradigm for rapid and guided discovery of materials through iterative combination of ML with HiTp experimentation.
  • Fig. 2 Performance and predictions of ML models for MG formation.

    (A) ROC curve for the model that predicts melt-spun GFA cross-validated against melt-spun observations from the LB handbook. (B) ROC curve from the melt-spun (blue curve) and stacked (pink curve) model cross-validated against observations of MGs in the LB handbook, synthesized by sputtering. Predictions of the GFL for compositions in Co-V-Zr synthesized by melt spinning (C) and sputter co-deposition (D).

  • Fig. 3 Comparison of the first-generation ML model and PCTs.

    (A) Prediction of high GFL based on the theories of Yang and Zhang (10). (B) Prediction of GFL based on efficient packing model (12). The white line is the ideal packing prediction, the purple band shows 1% deviation from ideal packing, and the green band shows 2% deviation from ideal packing. (C) Prediction of GFL from ML model with incorporation of the two PCTs (stacked + PCT).

  • Fig. 4 Comparison of new HiTp experimental results with the first-generation predictions.

    (A) Prediction of GFL for sputter co-deposition from ML model with incorporation of PCTs. (B) FWHM of the FSDP measured in HiTp XRD experiments. (C) Map of the GFR based on a glass formation threshold determined based on the FWHM of FSDP of amorphous silica (a-silica) in XRD measurements (see text for details).

  • Fig. 5 Higher-generation ML models.

    (A) Revised predictions for Co-V-Zr ternary. (B) Predictions of ternaries with the largest as-yet-unexplored GFR. (C) Comparison of the ROC curves for first-, second-, and third-generation models cross-validated against all available sputtered co-deposited synthesis data (LB + HiTp).

  • Fig. 6 Comparison of first- and second-generation predictions with HiTp experimental results for Co-Ti-Zr (first row), Co-Fe-Zr (second row), and Fe-Ti-Nb (third row) ternary.

    (A1 to A3) Prediction of GFL from the first-generation ML model. (B1 to B3) Revised predictions from the second-generation ML model. (C1 to C3) HiTp experimental map of the FWHM of the FSDP in XRD measurements. (D1 to D3) Experimental map of the GFR derived after application of the glass formation threshold based on amorphous silica applied to data in (C1) to (C3). Purple, glass; yellow, not glass.

Supplementary Materials

  • Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/4/4/eaaq1566/DC1

    ML models

    HiTp experiments

    Published data on Co-V-Zr

    Correlation between the liquidus surfaces, FWHM of FSDP, and GFR

    Estimation of the rate of success for discovering new MGs

    fig. S1. Performance of four approaches of including synthesis method into an ML model on a cross-validation test.

    fig. S2. Prediction of GFR by two physiochemical models.

    fig. S3. Elemental gun configuration on AJA ATC Orion 5 sputtering system.

    fig. S4. Accuracy of composition calculated by the in-house sputter model.

    fig. S5. Experimental determination of glass formation in Co-V-Zr ternary.

    fig. S6. Published reports of MGs in Co-V-Zr ternary.

    fig. S7. Correlation between the liquidus surfaces, FWHM of FSDP, and GFR.

    table S1. Sputtering parameters for all samples.

    References (38, 39)

  • Supplementary Materials

    This PDF file includes:

    • ML models
    • HiTp experiments
    • Published data on Co-V-Zr
    • Correlation between the liquidus surfaces, FWHM of FSDP, and GFR
    • Estimation of the rate of success for discovering new MGs
    • fig. S1. Performance of four approaches of including synthesis method into an ML model on a cross-validation test.
    • fig. S2. Prediction of GFR by two physiochemical models.
    • fig. S3. Elemental gun configuration on AJA ATC Orion 5 sputtering system.
    • fig. S4. Accuracy of composition calculated by the in-house sputter model.
    • fig. S5. Experimental determination of glass formation in Co-V-Zr ternary.
    • fig. S6. Published reports of MGs in Co-V-Zr ternary.
    • fig. S7. Correlation between the liquidus surfaces, FWHM of FSDP, and GFR.
    • table S1. Sputtering parameters for all samples.
    • References (38, 39)

    Download PDF

    Files in this Data Supplement:

Navigate This Article