A data-driven approach for optimising supplement allocation to individual lactating dairy cows in pasture-based systems

• Individualised feeding optimises concentrate use by leveraging cow variability. • Individualised feeding strategy optimises concentrate use for maximum milk yield. • Data-driven ML optimisation models offer a new approach to individualised feeding. • Optimised concentrate allocation increased total milk yield by ≈ 8 % in this study. With feed costs accounting for about 40-60 % of milk production expenses in Australia, efficient supplementary concentrate allocation is crucial for profitability. Despite an increase in concentrate use per cow over the past decade, the average milk yield response remains about 1 L per kg of dry matter concentrate. While machine learning and data-driven optimisation are widely utilised in sectors such as engineering, healthcare, and finance, their application in feed optimisation within dairy farming has not been extensively researched. This study aims to develop a machine learning-based method to optimise individual cow supplement allocation, using similar total daily concentrate with a tolerance range allowing for a 2-10 % decrease, to maximise milk yield. Data from a controlled field study involving 130 lactating Holstein-Friesians (32,504 records) were analysed. Sixteen machine learning algorithms were evaluated to predict milk yield based on concentrate allocation and available cattle data (days in milk, daily milk yield and liveweight, and parity number). The Random Forest (RF) model was the best performer, achieving an R² of 0.60 and RMSE of 4.20 L/cow/day. Then 7371 records from 81 cows over 91 days were employed to run the concentrate levels optimisation using the Dirichlet-Rescale (DRS) algorithm and Monte Carlo simulation. The RF model and SciPy optimisation determined optimal individual cow allocations (5-9 kg/cow.day -1 ). Implementing this resulted in a herd-level 8 % increase in daily milk yield. This study highlights the potential benefits of adopting data-driven algorithms for individualised dairy feed optimisation based on observed correlations within existing management practices. While results suggest improvement over conventional flat-rate methods, the study is limited by the nature of the dataset, and findings reflect associations under standard practice rather than experimental manipulation of concentrate levels, requiring validation through controlled field trials to confirm practical efficacy and economic impact in actual dairy farming contexts.