A Bayesian experimental autonomous researcher for mechanical design

See allHide authors and affiliations

Science Advances  10 Apr 2020:
Vol. 6, no. 15, eaaz1708
DOI: 10.1126/sciadv.aaz1708
  • Fig. 1 BEAR for studying the mechanics of additively manufactured components.

    (A) Experimental system composed of (i) five dual extruder fused deposition modeling (FDM) printers (M3, MakerGear), (ii) a six-axis robotic arm (UR5e, Universal Robotics), (iii) a scale (CP225D, Sartorius), and (iv) a universal testing machine (5965, Instron Inc.). (Photo credit: Aldair E. Gongora and Bowen Xu, Boston University). (B) Model “crossed barrel” family of parametric structures with two circular platforms that are held apart by a series of n hollow columns of outer radius r and thickness t and that are twisted with an angle θ. Force F and corresponding displacement D from the testing of (C) a crossed barrel that did not yield before ~5 kN (designated too strong), (D) a crossed barrel that failed in a brittle manner (designated “brittle”), and (E) a crossed barrel that exhibited appreciable strength after an initial yield point (designated “ductile”).

  • Fig. 2 Experimental exploration of the toughness of a family of parametric structures.

    (A) Overlaid F versus D curves for 240 samples printed with x = (n, θ, r, t) = (8,100 ° ,2 mm,1.05 mm). (B) Experimental toughness U versus component mass m for the samples shown in (A). Red line denotes a linear fit with a correlation coefficient of 0.71. (C) U versus m for 1800 samples taken in a grid across the entire parameter space. Marker shape denotes the category of mechanical response. (D) Predicted toughness Ugrid based on a Gaussian process regression (GPR) trained on the 1800 experimental data points evaluated at x versus average U(x). The red line has zero intercept and a slope of one as a guide to the eye. (E) Surface plot of Ugrid across the entire 4D parameter space with the discretization of the experimental grid represented as white circles in the top right panel.

  • Fig. 3 Simulated learning using BO.

    Distribution of experimental points when guided using (A) MV and (B) EI decision-making policies. The color gradient indicates the start and end of the campaign. Axis limits are the same as in Fig. 2E. Performance P versus experiment number i of simulated Bayesian campaigns with noise added to each simulated measurement drawn from a zero-mean Gaussian with (C) SD σ = 0.1 J and (D) σ = 5 J. EI- and MV-guided campaigns are benchmarked against PE and the average result of selecting 100 experiments using Latin hypercube sampling (LHS). Shaded regions correspond to the middle two quartiles of 100 simulated campaigns. The inset bar charts show the distribution in P at i = 100.

  • Fig. 4 Optimization of a family of mechanical structures using the BEAR.

    Computed P from six experimental campaigns carried out by the BEAR using (A) MV and (B) EI. (C) Average U measured from 10 samples of the best predicted structure from each of the six experimental campaigns and the best-performance structure predicted by the grid search. (D) Experimental optimization of U versus time T with ticks to the left of each bar denoting measurements taken before that time, ticks to the right denoting the 10 samples taken at the end of the campaign to evaluate the best predicted sample, and bars denoting the average measurement of the 10 samples. In (C) and (D), error bars correspond to SD. (E) Photographs overlaid on the F versus D curve corresponding to a structure printed with the best-performance design (12, 131°, 1.95 mm, 1.4 mm). (Photo credit: Aldair E. Gongora, Boston University).

Supplementary Materials

  • Supplementary Materials

    A Bayesian experimental autonomous researcher for mechanical design

    Aldair E. Gongora, Bowen Xu, Wyatt Perry, Chika Okoye, Patrick Riley, Kristofer G. Reyes, Elise F. Morgan, Keith A. Brown

    Download Supplement

    The PDF file includes:

    • Figs. S1 to S4
    • Legend for movie S1

    Other Supplementary Material for this manuscript includes the following:

    Files in this Data Supplement:

Stay Connected to Science Advances

Navigate This Article