This article presents a method for inverse design of multiring optical fibres using machine learning. It proposes fibre configurations expected to yield specified beam-shaping parameters and verifies selected off-grid configurations using independent OptiFiber simulations. The proposed method creates machine-learning surrogate models from numerical simulations performed in OptiFiber. It utilises differential evolution optimisation to predict and optimise the refractive index (RI) profiles of individual fibre layers for given beam parameters. The study reused a dataset comprising 721 valid three-layer optical-fibre configurations previously generated using OptiFiber and reported in [12]. For each fibre configuration, the output light beam was simulated, and the full width at half maximum (FWHM), beam flatness and edge steepness were calculated. As the direct prediction of fibre configuration from beam parameters is ambiguous, a surrogate model was employed for subsequent use. The feasibility of using several regression models to create a surrogate model was analysed, including ExtraTrees, Random Forest, XGBoost, LightGBM, CatBoost, support vector regression, multi-layer perceptron, ridge regression and the k-nearest neighbours method. Ensemble tree-based models yielded the best predictive performance, with the ExtraTrees model providing the highest overall accuracy and a high degree of agreement between predicted and simulated beam parameters. An ablation study demonstrated that prediction accuracy can be improved by extending the feature set used to train the surrogate model to include descriptors of the differences between the RI values of individual optical fibre layers. The employed inverse design procedure enabled continuous estimation of the optical fibre configuration, reducing the need for exhaustive searches based on time-consuming simulations. The final inverse-design model was independently validated using 30 new OptiFiber simulations corresponding to continuous refractive-index configurations that were not present in the original simulation grid. For these off-grid cases, the ExtraTrees surrogate achieved R² values of 0.9982, 0.9922, and 0.9912 for FWHM, beam flatness, and edge steepness, respectively. The results demonstrate that the surrogate-assisted optimisation can generate previously unseen off-grid fibre configurations whose beam parameters are reproduced with high accuracy in independent OptiFiber simulations. The conclusions are restricted to interpolation within the investigated RI range of 1.447–1.455.