Abstract
Rigid polyurethane foam (RPUF) is highly flammable and releases substantial heat and smoke during combustion, highlighting the need for effective flame-retardant strategies. This thesis combines the experimental synthesis of flame retardants with interpretable machine learning (ML) modeling to enhance the flame retardancy of RPUF and reduce reliance on trial-and-error formulation design. Reactive nitrogen- and phosphorus-containing flame-retardant polyols (FRPN and FRPP) were synthesized and combined with 2-methyl-1,3-propanediol (MPD). An FRPN:FRPP ratio of 2:1 achieved a limiting oxygen index (LOI) of 27.2% and a UL-94 V-0 rating. However, FRPP alone increased smoke generation. To address this limitation, a boron-containing flame-retardant polyol (FRPB) was introduced to replace FRP<sub>P</sub>, forming a nitrogen–boron synergistic system. This system (FRPN:FRPB = 1:1) maintained mechanical integrity, achieved an LOI of 26.8%, and substantially reduced total heat release (−68.8%) and smoke production (−62.7%). </p><p>To further overcome the limitations of experimental trial-and-error in flame retardant design, an ML model was developed to predict the performance of formulations containing up to two reactive flame retardants, using 435 literature-sourced formulations (coefficient of determination, R² = 0.84). SHapley Additive exPlanations (SHAP) analysis guided the design of a phosphorus–boron flame retardant (THPO-B). When combined with FRPN (1:1, total loading 26.0 wt%), the system achieved a relative LOI prediction error of only 2.2%. Building on this framework, a generalized LOI prediction and mechanistic interpretation model was established for both reactive and additive flame retardants. The model, trained on 1511 formulations without relying on molecular descriptors, demonstrated high predictive accuracy (R² = 0.87) and broad applicability. The model provided practical dosage guidance, identified effective multi-mechanism combinations, and was experimentally validated with prediction errors below 5%, confirming its accuracy and transferability. Overall, this thesis establishes an integrated experimental–computational framework for the rational design of flame-retardant RPUF, offering high predictive accuracy and broad applicability to different flame-retardant types and formulation conditions.