Machine-Learning Prediction of Marshall Mix-Design Properties and Mechanical Characterization of DCR-Treated Asphaltic Concrete

Marshall mix design remains the dominant empirical framework for proportioning hot-mix asphalt, and coconut-derived by-products such as Desiccated Coconut Residue (DCR) have attracted interest as sustainable bitumen modifiers. Determining the optimum bitumen content (OBC) and characterizing mechanical performance is laboratory-intensive, which motivates data-driven prediction. This study evaluated whether regression models can predict six Marshall responses for DCR-modified asphalt concrete and characterized the mechanical performance of the modified mixtures at their respective OBC. Thirty averaged Marshall records covering bitumen contents of 4–6 %, two gradations (AC14, AC20), and three treatments (Control, untreated DCR, treated DCR) were modeled using eight algorithms and a quadratic response-surface baseline. Models were assessed by pooled out-of-fold prediction under random K-fold, leave-one-mixture-out, and leave-one-bitumen-level-out validation. For the AC14 mixtures, OBC values were 5.30% (Control), 4.95% (untreated), and 5.10% (treated); resilient modulus, indirect tensile strength (ITS), tensile strength ratio (TSR), and dynamic creep were measured for each mixture prepared at its own OBC. Pooled out-of-fold R² reached 0.89 for voids filled with bitumen, 0.88 for air voids, 0.83 for bulk density, 0.81 for voids in mineral aggregate, 0.53 for flow, and 0.49 for stability. The quadratic response surface matched or exceeded the machine-learning models on four of six responses. Under leave-one-mixture-out validation, R² became negative for most responses, whereas leave-one-bitumen-level-out validation retained R² of 0.47–0.91. The models are therefore interpolation tools within the tested design space rather than general predictors. The untreated mixture required the lowest binder content, and the treated mixture recorded the highest dry ITS (1,148.7 kPa) and TSR (50.3%), although all mixtures fell below common TSR acceptance limits despite the hydrated-lime filler. ABSTRAK: Reka bentuk campuran Marshall, kekal sebagai kerangka empirik utama bagi menentukan kadar campuran asfalt panas, dan hasil sampingan berasaskan kelapa seperti Sisa Kelapa Kering (DCR) telah menarik minat pengubah suai bitumen yang mampan. Penentuan kandungan bitumen optimum (OBC) dan pencirian prestasi mekanikal memerlukan kerja makmal yang intensif, sekaligus mendorong penggunaan ramalan berasaskan data. Kajian ini menilai sama ada model regresi mampu meramal enam tindak balas Marshall bagi konkrit asfalt terubah suai DCR, serta mencirikan prestasi mekanikal campuran tersebut pada OBC masing-masing. Sebanyak 30 rekod Marshall purata merangkumi kandungan bitumen 4–6 %, dua gradasi (AC14, AC20) dan tiga rawatan (Kawalan, DCR tidak dirawat, dan DCR dirawat) dimodelkan menggunakan lapan algoritma serta model permukaan sambutan kuadratik sebagai asas perbandingan. Model dinilai melalui ramalan terkumpul luar-lipatan di bawah pengesahan K-lipatan rawak, tinggal-satu-campuran-keluar dan tinggal-satu-aras-bitumen-keluar. Bagi campuran AC14, nilai OBC ialah 5.30 % (Kawalan), 4.95 % (tidak dirawat) dan 5.10 % (dirawat). Modulus kenyal, kekuatan tegangan tak langsung (ITS), nisbah kekuatan tegangan (TSR) dan rayapan dinamik diukur dengan setiap campuran yang disediakan pada OBC tersendiri. R² terkumpul luar-lipatan mencapai 0.89 bagi lompang terisi bitumen, 0.88 bagi lompang udara, 0.83 bagi ketumpatan pukal, 0.81 bagi lompang dalam agregat mineral, 0.53 bagi aliran dan 0.49 bagi kestabilan. Model permukaan sambutan kuadratik menyamai atau mengatasi model pembelajaran mesin bagi empat daripada enam tindak balas. Di bawah pengesahan tinggal-satu-campuran-keluar, R² menjadi negatif bagi kebanyakan tindak balas, manakala pengesahan tinggal-satu-aras-bitumen-keluar mengekalkan R² 0.47–0.91. Justeru model ini merupakan alat interpolasi dalam ruang reka bentuk yang diuji dan bukan peramal umum. Campuran tidak dirawat memerlukan kandungan pengikat terendah, manakala campuran dirawat mencatat ITS kering tertinggi (1,148.7 kPa) dan TSR tertinggi (50.3 %), walaupun semua campuran berada di bawah had penerimaan TSR lazim.