Surrogate-Based Sequential Experimental Design of Metallurgical Solid-Waste-Based Gel Materials for Cemented Paste Backfill

Cemented paste backfill (CPB) provides a pathway for utilizing metallurgical solid wastes, in which blended precursor constituents hydrate and develop strength-bearing gel products within the hardened backfill. Screening multi-component formulations is experimentally intensive because each candidate requires preparation, curing, and multi-age strength testing. The primary contribution of this study is a data-driven virtual experimental environment that integrates multi-age UCS surrogates, constrained chemistry-to-recipe mapping, cost and feasibility evaluation, and Markovian campaign updates. This environment is used to investigate sequential experimental-design strategies for balancing strength, cost, and recipe feasibility. Forty-eight measured mixtures comprising five gel-forming precursor constituents, four process variables, and 3-, 7-, and 28-day uniaxial compressive strength (UCS) were used to construct age-specific surrogate models. A chemistry-space representation based on six major oxides was introduced, and candidate chemistries were mapped to executable formulations through constrained optimization. Sequential formulation design was cast as a finite-horizon Markov decision process and investigated using reinforcement learning (RL) and query-by-committee active learning (AL). Each campaign comprised 50 surrogate-generated virtual queries; no additional specimens were prepared for these campaign evaluations. Under normal supply, LSTM-based RL obtained the highest five-seed mean dimensionless strength-cost score (0.1895), equivalent to 84.8% of the optimistic PSO reference; the best AL configuration reached 0.1753. A supplementary 50-seed virtual comparison yielded 0.1818 versus 0.1729 (paired t-test, p = 0.0064). AL and RL were statistically unresolved when steel slag was unavailable, whereas AL retained a higher fraction of the case-specific PSO reference after the EAF reduction-slag price doubled. Randomized feed-chemistry simulations further showed that the workflow can re-project target chemistries and identify feasible candidates within the domain represented by the surrogate data. The resulting computational environment supports efficient screening of sequential design strategies and prioritization of promising precursor recipes for subsequent laboratory preparation, strength verification, and hydrate-gel characterization.