The use of carbon fiber-reinforced polymer (CFRP) tubes as crash boxes has become a subject of interest due to their high specific strength and energy absorption capabilities. This study investigates the crashworthiness performance of rectangular tubes made of CFRP, with and without holes and polyurethane foam (PUF)-filled inner structures. The designed tubes were subjected to quasi-static axial compression loading. In addition to carefully documenting failure histories, data on crash load and displacement responses were methodically recorded during testing. To evaluate crashworthiness performance, three design parameters were considered: hole diameter, the number of holes in both the x and y directions, and whether the tube was filled with foam or left unfilled. Machine learning (ML) was also used to reduce the time and cost by predicting the crashworthiness indicators of the tubes from fewer experiments. A collection of ML algorithms such as decision tree regressor (DTR), linear regressor (LR), ridge regressor (RR), lasso regressor (LAR), elastic nets (ENs), and multi-layer perceptron (MLP) have been utilized to predict crashworthiness indicators such as initial peak force (Pip), mean crushing force (Pm) and energy absorption (EA) of the design tubes from the experimental data. The experimental results showed that PUF-filling significantly enhanced crashworthiness properties, with Pm and EA increasing by nearly threefold compared to unfilled tubes. Furthermore, in unfilled tubes, the introduction of holes led to varying effects depending on the hole diameter and placement. Meanwhile, in PUF-filled tubes, the presence of holes reduced the crashworthiness performance. For ML prediction, the DTR achieved the best accuracy with the lowest value of root mean squared error (RMSE) and mean absolute percentage error (MAPE) of 1251 and 11.37%, respectively. These findings demonstrate both the importance of PUF-filled, perforation configurations and the feasibility of ML models in optimizing CFRP crash box designs.