Since the beginnings of commerce, marine terminals have been epicenters of logistics. But, the world around them has changed immeasurably. The ways of conducting business within ports are primed for the next evolutionary step. Machine learning is the decisive tool that enables changes with profound impacts for global logistics.

To realize this potential, a TCS team of machine-learning scientists at the TCS Pace Port* in New York collaborated with the team behind TCS DynaPORT*, a leading Terminal Operating System (TOS) globally deployed in over 80 ports. Since it's at the center of task management and flow of data for nearly all port activities, the TOS is the decisive system for harnessing AI to impact all operations. However, the wide array of functionalities conducted in a complex and fast-paced port ecosystem cannot easily achieve the full potential of this in one swift transition.

Success for the joint team meant selecting the TOS functionality best suited for using machine-learning solutions. It also involved using a platform that used cutting-edge computation tools to produce the best results. The Intel^® oneAPI platform proved to be the perfect environment where the value of AI can be unlocked. For a port ecosystem where moving cargo is a frequent bottleneck, scheduling trucks is perfectly suited for that application. 

The results for training the model on a GPU with X^e Architecture proved fast enough to keep up with the rate of innovation. Changing a single line of code moved the model over from the GPU to run inference on a CPU. Optimized performance on both architectures through oneAPI resulted with a 30% reduction in overall development. This is from using the same processors without the unified runtime that oneAPI provided.

With development and implementation finished, analysis of the final turnaround time (TAT) for the prediction model’s accuracy produced desirable results showing the possibility to be used in live scenarios at seaports. Predictions saw accuracy of  around eight minutes of the actual results, which was compelling given that actual TAT varied from 10 to 120 minutes. This level of accuracy, even from the initial data collected, provides a valuable tool enabling haulers to plan their fleet optimally while terminals can induce traffic behavior favoring optimal asset use. Designed to improve as more data is gathered, this neural network also creates growing value. It learns from more data in the environments where the network is deployed to further integrate it with a smartport ecosystem.