Predicting the performance of a functional ecological substrate via a generative model based on an orthogonal experiment

Ecological slope protection plays a crucial role in increasing slope stability and preventing landslide disasters. The effectiveness of ecological slope protection is influenced by the ecological substrate used. In this study, the impact of materials and physical properties on the performance of ecological substrates based on red soil was investigated. Raw materials such as red soil, organic fertilizer, cement, sawdust, and ferrous sulfate were utilized for the ecological substrate, with Cynodon dactylon selected as the slope-protection plant. Orthogonal experiments L25(5³) with three factors and five levels were conducted to analyze the growth status of Cynodon dactylon, shear strength, and degree of disintegration after dry–wet cycles. The study examined the interrelationships among multiple factors for ecological substrates using Pearson correlation and statistical significance tests. Range analysis was applied to determine the optimal formula. Furthermore, an ecological substrate neural network (ESNN) model based on a variational autoencoder (VAE) algorithm was initially designed to verify the optimal formula of ecological substrates and predict the performance of experiment arrays generated by VAE. The research findings indicated that the optimal formula obtained through ESNN model closely aligned with that obtained by range analysis. The excellent prediction performance of ESNN model was confirmed by the field experiments with random level values assigned to three factors. Therefore, the ESNN model can objectively evaluate orthogonal experiments' performance and provide property predictions for plant growth and slope stability for on-site construction conditions. This research provides a design method suitable for road ecological slope protection under red soil conditions.