Harbin Institute of Technology Builds AI Curriculum Inspired by Cornell’s ECE4950
In a bold stride toward redefining engineering education in the age of artificial intelligence, researchers at Harbin Institute of Technology (HIT) have developed a novel AI-integrated curriculum framework modeled after Cornell University’s renowned ECE4950 course. Spearheaded by Liguo Wang and Ben Guo from HIT’s School of Electrical Engineering and Automation, alongside Li Liu from the School of Chemical Engineering, the initiative represents a strategic fusion of foundational theory, hardware design, algorithm development, simulation, and hands-on experimentation—all tailored to cultivate a new generation of innovation leaders equipped for global technological challenges.
The curriculum, detailed in a recent paper published in Experiment Science and Technology, outlines a comprehensive educational architecture that bridges classical engineering disciplines with modern machine learning techniques, particularly leveraging Gradient Boosting Decision Trees (GBDT). By anchoring the learning journey in real-world industrial applications—specifically, intelligent temperature control in oil pipeline heating systems—the team has created a pedagogical model that emphasizes not just technical proficiency, but systems thinking, cross-disciplinary integration, and enterprise-ready problem-solving.
This approach arrives at a critical juncture. As nations worldwide accelerate investments in AI infrastructure and talent development, China has intensified its focus on producing engineers who can lead in both innovation and implementation. The 2018 national directive from the Ministry of Education calling for “AI + Smart Learning” educational ecosystems provided the policy impetus, but translating that vision into classroom practice demanded more than syllabus tweaks—it required a fundamental reimagining of how engineering knowledge is structured, sequenced, and applied.
Wang and his colleagues found inspiration during academic exchanges at Cornell University, where ECE4950—officially titled “Special Topics in Electrical and Computer Engineering”—has long served as a capstone experience that immerses students in open-ended, industry-relevant engineering problems. Unlike conventional courses that silo theory from practice, ECE4950 integrates mathematical modeling, system design, computational simulation, and hardware prototyping within a single project lifecycle. This holistic methodology resonated deeply with the HIT team, who saw in it a blueprint for nurturing the kind of adaptive, systems-oriented engineers China urgently needs.
“What distinguishes ECE4950 is its insistence on starting with the physics of a problem,” explained Liguo Wang. “You don’t jump to coding or circuit design. You first understand the underlying mechanisms—why a system behaves the way it does. Only then do you build the tools to control or optimize it. That sequence—mechanism analysis → system design → algorithm development → simulation → experimental validation—is now the backbone of our AI curriculum.”
To operationalize this vision, the team mapped existing university courses onto a five-stage AI engineering pipeline. At the foundational level, general education courses like Calculus, Electromagnetism, and University Physics are no longer treated as abstract prerequisites. Instead, they are explicitly linked to “mechanism analysis”—the process of deriving differential equations that describe physical phenomena, analyzing electromagnetic coupling in conductive materials, or uncovering thermodynamic principles governing heat transfer in pipelines.
This grounding in first principles ensures that students approach AI not as a black box, but as a tool to augment deep domain understanding. For instance, in the oil pipeline heating case study, students use calculus to model thermal diffusion and electromagnetism to compute skin depth—the critical parameter that determines how deeply alternating current penetrates a metal pipe and generates heat. These physical insights directly inform the selection of optimal operating frequencies, a decision that would otherwise be arbitrary if treated purely as a data-driven optimization problem.
The second stage—system design—draws on core electrical engineering coursework: Circuit Theory, Analog Electronics, and Digital Electronics. Here, students translate theoretical models into functional hardware. Using Thevenin and Norton theorems, they analyze transient and steady-state responses of power circuits. They design signal conditioning modules with operational amplifiers like OP07 and LM324, develop gate driver circuits for IGBTs using Darlington transistors such as the 2N3055, and architect digital control systems around the Texas Instruments DSP28335 microcontroller. Crucially, the curriculum emphasizes real-world constraints: electromagnetic interference (EMI), load interactions between circuit stages, and signal integrity—issues often glossed over in purely software-centric AI programs.
This hardware-aware foundation sets the stage for the third and most distinctive phase: algorithm development with GBDT. Rather than defaulting to deep learning—a common but often overkill approach for industrial control problems—the HIT team selected GBDT for its interpretability, efficiency, and strong performance on tabular data. Students learn to frame engineering objectives (e.g., minimizing energy consumption while maintaining pipeline temperature at 40°C) as regression tasks. Using field-collected data on pipe diameter, current, frequency, and temperature, the GBDT model predicts the optimal heating frequency. This prediction is then embedded within a C-language control framework running on the DSP28335, closing the loop between data science and real-time hardware execution.
The integration of Automatic Control Theory further elevates this phase. Students apply classical control concepts—such as Nyquist stability criteria—to tune PI controllers that work in tandem with the GBDT algorithm. This hybrid approach combines the robustness of model-based control with the adaptability of machine learning, reflecting the pragmatic engineering ethos that dominates industrial AI deployments.
Simulation forms the fourth pillar. Leveraging industry-standard tools like MATLAB/Simulink, PSCAD, and ANSYS, students validate their designs before physical prototyping. MATLAB models predict control performance under various disturbances; PSCAD simulates power electronics behavior, including EMI effects; ANSYS performs finite element analysis to visualize current density distribution and thermal profiles in the pipeline. These simulations not only reduce trial-and-error costs but also teach students to define realistic boundary conditions and interpret multi-physics interactions—skills essential for complex system engineering.
Finally, experimental validation grounds the entire process in empirical reality. Using standard lab instruments—multimeters, oscilloscopes, and the HIOKI 3198 power analyzer—students measure voltage, current waveforms, harmonic distortion, power factor, and energy consumption over extended periods. In the pipeline heating project, these measurements confirmed that the AI-optimized system achieved the target temperature of 40°C at 800 Hz with significantly lower energy use compared to conventional fixed-frequency heaters. More importantly, the validation step reinforces the scientific method: hypotheses derived from theory and simulation must withstand real-world scrutiny.
The curriculum’s effectiveness has already shown promising early results. Since its pilot launch in June 2018, a cohort of six students completed the 12-hour intensive module. Two have since joined Eaton Corporation’s AI research division in the United States—a testament to the program’s alignment with global industry expectations. The collaboration with Eaton, formalized through the HIT-Eaton Engineering Practice Education Center, provides students with access to industrial-grade equipment and real engineering challenges, blurring the line between academia and enterprise.
Critically, the HIT model addresses a growing concern in AI education: the “skills gap” between data scientists and domain engineers. Many AI programs produce graduates fluent in Python and neural networks but ill-equipped to interface with physical systems, power electronics, or safety-critical infrastructure. Conversely, traditional engineering programs often lag in computational literacy. By weaving AI into the fabric of electrical engineering—from Maxwell’s equations to microcontroller firmware—Wang’s team ensures graduates speak both languages.
This integrative philosophy also reflects broader shifts in how AI is deployed in industry. While consumer tech dominates headlines with large language models and generative AI, the industrial sector relies heavily on “narrow AI” solutions—targeted, efficient algorithms that solve specific operational problems. Predictive maintenance, energy optimization, quality control, and adaptive control loops are where AI delivers tangible ROI in manufacturing, energy, and transportation. The HIT curriculum prepares students precisely for this reality.
Moreover, the emphasis on interpretability and physics-informed AI aligns with emerging regulatory and ethical standards. As AI systems increasingly manage critical infrastructure, black-box models become unacceptable. Engineers must understand why an algorithm makes a decision, especially when that decision affects safety or efficiency. By rooting AI development in physical laws and validating every step through simulation and experiment, the HIT approach builds inherently more trustworthy systems.
Looking ahead, the team plans to expand the framework to other domains—renewable energy integration, smart grids, electric vehicle charging systems—while maintaining the core pedagogical sequence. They also aim to open-source course materials and collaborate with other universities to establish a benchmark for AI-infused engineering education in China and beyond.
In an era where technological leadership hinges on the ability to educate adaptable, systems-thinking engineers, the work by Wang, Liu, and Guo offers a compelling template. It proves that AI education need not abandon engineering fundamentals; rather, it can revitalize them. By standing on the shoulders of classical disciplines while reaching toward the frontiers of machine learning, Harbin Institute of Technology is not just teaching AI—it’s redefining what it means to be an engineer in the 21st century.
Authors: Liguo Wang¹, Li Liu², Ben Guo¹
Affiliations:
¹ School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin 150001, China
² School of Chemical Engineering, Harbin Institute of Technology, Harbin 150001, China
Journal: Experiment Science and Technology
DOI: 10.12179/1672-4550.20200177