Machine Learning Enables Rapid Screening of Reactive Fly Ashes Based on Their Network Topology

Fly ash, a byproduct of coal combustion, can be used as supplementary cementitious material (SCM) to replace ordinary portland cement (OPC) in concrete. This generates revenue for coal power plant operators and also reduces the CO2 intensity of the binder fraction of a concrete (each ton of OPC replaced by fly ash results in 0.9 ton of avoided CO2 emissions, if the fly ash is considered to have no carbon footprint). However, the use of fly ash in concrete has thus far been limited to replacement levels less than 20 mass % due to uncertainties in their performance as SCM. Although the ability of a fly ash to replace cement in concrete is largely determined by the reactivity of its amorphous phase, characterizing fly ashes' amorphous phase is complex and cost prohibitive, which has thus far prevented any high-throughput screening of fly ashes to assess their suitability as SCMs. Here, we introduce a machine-learning-based methodology that enables robust screening of reactive fly ashes based solely on fast, inexpensive bulk characterization (X-ray fluorescence: XRF), by using the network topology of fly ashes' amorphous phase as a structural proxy for their reactivity. On the basis of a data set of more than 100 fly ashes, we train an artificial neural network (ANN) model that offers accurate predictions of the mass fraction of fly ashes' amorphous phase and the network topology thereof. This new method seeks to maximize the beneficial use of fly ashes obtained from routine production, as well as to identify opportunities for the reclamation of ashes that are presently stored in impoundments.