Skip to content

Interactive Figures

This page embeds interactive versions of the key figures from the paper "ML-guided screening of chalcogenide perovskites as solar energy materials".

The figures are generated by running:

python docs/generate_interactive_figures.py

from the repository root. Output HTML files land in docs/assets/figures/ (CDN-linked Plotly, requires an internet connection to render).

Figure 2 — Experimental data distribution of ABX₃ compounds

Distribution of tolerance factors and logistic-calibrated probability P(τ*) for the training + test set of ABX₃ chalcogenide perovskites.

τ* (SISSO-derived) distribution

t* (Jess et al.) distribution

t* vs P(τ*)

Calibrated probability P(τ*) as a function of the Jess et al. tolerance factor. The green band marks the stable region (0.84 ≤ t* ≤ 1.02).

Figure 3 — τ* stability over ionic-radii space

Logistic-calibrated probability P(τ*) of perovskite-type structural stability as a function of the A-site (r_A) and B-site (r_B) ionic radii. Hover over the experimental compound markers to see the formula, radii, and stability class.

ABS₃

ABSe₃

Show raw τ* values instead of P(τ*) **ABS₃** **ABSe₃**

Figure 4 — Crystal structures (polyhedral view)

Interactive 3D structure CrystaLLM-predicted perovskite structures. BX₆ octahedra are rendered as semi-transparent polyhedra. Drag to rotate, scroll to zoom, hover atoms for coordinates.

Select a compound from the dropdown to load its structure:

Figure S3 — Structure relaxation: initial vs. relaxed energy

Scatter plot comparing the initial (CrystaLLM-generated) and UMA-relaxed energy per atom for each ABX₃ candidate. Points are coloured by the energy reduction ΔE = Einit − Erelax per atom (green = small change, red = large reduction). The dashed diagonal line marks the no-change baseline (Erelax = Einit). Hover to see the compound name and exact energy values.

Figure 5 — CrabNet-estimated bandgaps (element matrix)

Element–element matrix for ABS₃ and ABSe₃ compositions. Color encodes the CrabNet-predicted bandgap. Squares with a black border correspond to compositions for which CrystaLLM generated a corner-sharing perovskite-type structure. Hover to see the formula and exact bandgap.

ABS₃ bandgap matrix

ABSe₃ bandgap matrix

Figure S2 — P(τ*) stability probability (element matrix)

Same element–element layout, now coloured by the logistic-calibrated probability of perovskite-type stability derived from τ*.

ABS₃ probability matrix

ABSe₃ probability matrix

Figure 6 — Pareto front: Supply Risk vs bandgap deviation

Multi-objective evaluation of material sustainability (Supply Risk, SR) versus deviation of CrabNet-predicted bandgap from the optoelectronic optimum.

  • Red triangles — Pareto-optimal candidates (minimum SR and minimum |E_g^ − E_g^opt|)
  • Blue squares — materials within 10 % of both minima
  • Gray circles — remaining candidates

Hover over any point to see the formula, bandgap, and SR.

Single-junction photovoltaics (E_g^opt = 1.34 eV)

Tandem top cell (E_g^opt = 1.71 eV)

Figure 7 — Pareto front: Crystal-likeness Score (CLS) vs SR

Three-objective Pareto front combining bandgap deviation, Supply Risk, and 1 − CLS (lower = more synthesisable). Triangles indicate single-junction Pareto-optimal materials; squares indicate tandem-optimal materials.

Single-junction (E_g^opt = 1.34 eV)

Tandem top cell (E_g^opt = 1.71 eV)

Figure S9 — Spearman rank correlation matrix

Pairwise Spearman rank correlations (ρ) between the four screening metrics for the CrystaLLM-predicted candidate set (n = 54). All |ρ| ≤ 0.27, confirming that each pipeline stage captures largely independent information. Hover for exact values.