The 12,700× Speedup That Could Make Electric Vehicles More Efficient

The Science

Every electric vehicle on the road today faces a fundamental engineering tension: the电机 that drives it must switch between electrical states thousands of times per second, and each switch costs energy. The faster the switching, the smoother the motor runs—but the more power dissipates as heat in the transistors doing the switching. It's a tradeoff engineers have lived with for decades, usually accepting that perfect smoothness requires punishing computational costs.

A team from the University of Michigan has now found a way to get nearly optimal motor control without paying that price. Their method, published in a new paper by Huu-Thinh Do, Trung B. Tran, Jing Sun, and Ilya Kolmanovsky, achieves performance comparable to the most sophisticated existing approach—finite-control-set model predictive control, or FCS-MPC—while requiring far less computing power to run. The key lies in a mathematical technique borrowed from convex optimization called Linear Matrix Inequalities, or LMIs, which allows the researchers to pre-compute the hardest part of the control problem offline, then apply it in real-time with minimal effort.

The motors they're targeting are called permanent-magnet synchronous motors—PMSMs in the literature—and they're the dominant technology in electric vehicle drivetrains today. A PMSM generates torque through the interaction between permanent magnets embedded in its rotor and magnetic fields produced by coils in its surrounding stator. To control which way the motor turns and how fast, you need to manage the voltages applied to those coils. In a typical setup, this is done through a voltage-source inverter: a bank of transistors that switches between discrete voltage levels to produce the desired magnetic fields.

The challenge is that the inverter can only produce a finite number of voltage states—eight for a two-level inverter, twenty-seven for a three-level one. Choosing which state to apply at each moment is a discrete optimization problem, and optimizing over many time steps into the future makes it computationally intense. FCS-MPC handles this by solving a constrained optimization at each control cycle, testing all possible switching sequences and picking the best one. The results are excellent, but the computational burden scales poorly with the number of voltage levels and the prediction horizon length. For a three-level inverter with a horizon of three steps, that's 27³ = 19,683 possible sequences to evaluate every control cycle—at 50,000 times per second. Even on modern hardware, that's demanding.

The Michigan team took a different approach. Instead of solving the full optimization online, they approximate the solution to a related mathematical object called the value function. In optimal control theory, the value function captures the long-term cost of starting from any given state: if you're here, with this current flowing through those coils, how much tracking error will you accumulate over infinite time if you make the best possible choices going forward? Knowing this function completely would give you the perfect controller—you'd simply pick the action that leads to the lowest future cost. But computing it exactly for a nonlinear system with discrete switching inputs is intractable.

What Do and colleagues did was approximate this function using a quadratic form—essentially, a weighted sum of squared errors. They constrained the approximation to be independent of the motor's rotational angle, which seems counterintuitive but turns out to work well in practice because the angle is an uncontrolled variable. The genius move was reformulating the approximation problem as a set of Linear Matrix Inequalities. LMIs are a class of convex constraints that can be solved efficiently using semidefinite programming, even when they involve infinitely many conditions. By parameterizing the value function quadratically and enforcing a relaxed version of the Bellman inequality—the fundamental equation of dynamic programming—the researchers converted the infinite-horizon optimal control problem into a tractable offline optimization.

The offline computation produces a function that underestimates the true optimal cost, which is actually useful: using an underestimate as the tail cost in a horizon-one controller guarantees the closed-loop system will perform at least as well as the approximation predicts. In practice, this means the pre-computed value function acts like a learned "cost-to-go" map. When the controller encounters a particular current state, it evaluates the predicted cost of each possible voltage vector using this map and picks the cheapest option. No tree search, no horizon extension, just a single-step minimization using the pre-computed function.

The researchers implemented this scheme—calling it LMI-ADP, for LMI-based approximate dynamic programming—and tested it on a detailed motor model parameterized for a 15,000 RPM machine with parameters extracted from finite-element analysis software. They synthesized value functions for inverters with two through five voltage levels, then evaluated closed-loop performance in simulation against two benchmarks: the indirect method with space vector modulation (PI-SVM), which is the industry standard, and FCS-MPC with a three-step prediction horizon, which represents the current state of the art in academic research.

What They Found

The results are striking in their implications for the tradeoff between complexity and performance.

The LMI-ADP controller achieves an average current ripple of approximately 10.8 A at a switching frequency around 3.3 kHz when operating a two-level inverter. FCS-MPC under the same conditions produces about 9.1 A of ripple at 4.0 kHz. The indirect method (PI-SVM) performs worst on both metrics: roughly 12.7 A of ripple at 5.0 kHz switching frequency. In raw numbers, LMI-ADP sits between the two—about 19% worse than FCS-MPC on ripple, but 18% better than PI-SVM. The switching frequencies tell a similar story: LMI-ADP switches less often than either benchmark, which translates directly to lower conduction losses in the inverter hardware.

The trade-off curves in the paper make this relationship viscerally clear. When plotted as switching frequency versus current ripple, the three methods occupy distinct but overlapping regions. PI-SVM occupies the high-switching, high-ripple corner: it switches constantly at whatever rate the modulation scheme demands, regardless of whether the motor actually needs correction. FCS-MPC wanders around the middle, finding sequences that minimize a weighted combination of ripple and switching, but constrained by the computational cost of evaluating many candidate sequences. LMI-ADP traces a curve that partially overlaps with FCS-MPC, particularly at lower switching frequencies, but extends further into the low-switching region—achieving ripple levels FCS-MPC can only match with significantly higher switching rates.

This is the core empirical finding: by pre-computing a good approximation to the optimal value function, LMI-ADP can achieve ripple performance that FCS-MPC only matches when it switches substantially more often. The reason, according to the researchers, is that the pre-computed function implicitly contains information about the infinite-horizon optimal policy. When FCS-MPC looks only three steps ahead, it can't see the long-term consequences of its choices. LMI-ADP's horizon-one controller, augmented with the value function approximation, effectively reasons about the future without doing the costly search.

The computational advantage is equally dramatic. The paper reports that offline function generation takes roughly 3.4 seconds on a standard laptop for the two-level inverter case, compared to 12 hours for value iteration on a comparable grid. This is a factor of roughly 12,700× speedup in the offline phase. More importantly, the online computational burden per control step is essentially negligible: evaluating the pre-computed quadratic function for each of the p possible voltage vectors requires only matrix-vector multiplications, which modern microcontrollers can perform in microseconds. The paper doesn't explicitly state online solve times, but the structure of the computation—evaluate p quadratic forms and pick the minimum—suggests it falls well within the 20-microsecond sampling period at 50 kHz.

The stability analysis yields additional insight. The researchers certify that the closed-loop system remains bounded in a region they call Ω_δ, which is defined by the sublevel set of the approximate value function. Within this region, they can guarantee that the average current ripple over infinite time is bounded by a computable quantity ν. For the two-level inverter case, ν is approximately 140, meaning the time-averaged squared current error is guaranteed to stay below this level. While this bound is conservative—it arises from sufficient conditions that are stronger than necessary—it provides a formal guarantee that the method won't cause the motor to diverge or exhibit unbounded oscillations.

The vector field visualization in Figure 2a shows something interesting: the closed-loop trajectories form distinct regions of attraction corresponding to each of the eight possible voltage vectors. The boundaries between these regions are piecewise linear functions of the state, which is exactly what Lemma 3.2 predicts. At θ_re = 0, the partition of the state space into decision regions takes on a characteristic pattern, with each region assigned to one inverter state. As the rotational angle varies over a full electrical cycle, these regions rotate and distort, but the qualitative structure remains—each voltage vector "owns" a wedge of the state space.

Why This Changes Things

The electric vehicle industry is at an inflection point. Battery costs have fallen far enough that the motor and power electronics are now among the remaining opportunities for efficiency gains. A typical EV motor controller switches at 10-20 kHz, and the associated losses in the inverter can account for 2-5% of the drivetrain's energy losses. Reducing switching frequency while maintaining torque smoothness directly improves range—a precious commodity in a market where 300 miles feels inadequate.

Current production vehicles use variations of the indirect control method (space vector modulation with PI regulators) because it's computationally cheap and robust. The academic community has spent two decades developing FCS-MPC as the theoretically optimal replacement, but adoption has been limited by the computational requirements. A three-level inverter with a three-step horizon demands real-time optimization with 27³ = 19,683 candidate sequences. Automotive microcontrollers can handle this, but the power budget is nontrivial, and the worst-case execution time can be difficult to guarantee—important when you're certifying safety-critical systems.

LMI-ADP offers a third path: the expensive computation happens once, offline, on an engineer's workstation. The result is a few kilobytes of matrix coefficients. Applying those coefficients in real-time requires only a handful of multiplications and additions. For a two-level inverter with eight voltage vectors, that's eight quadratic function evaluations per control step—trivial even for a modest processor. For a five-level inverter with 125 vectors, it's still manageable: 125 quadratic evaluations at 50 kHz is well within the capability of a Cortex-M4 class microcontroller.

The implications extend beyond electric vehicles. Permanent-magnet synchronous motors are ubiquitous in industrial robotics, aerospace actuators, HVAC systems, and consumer electronics. Anywhere motor efficiency matters—and in a world of increasing electrification, that's everywhere—this approach offers a way to implement sophisticated optimal control without the computational overhead.

There's also a systems-level benefit that's easy to overlook. When controllers require heavy real-time computation, engineers must spec processors with margin for worst-case loads, leading to overdesigned hardware. When the computation is lightweight, you can use cheaper processors, or run the control loop faster for better bandwidth, or add diagnostic monitoring without sacrificing headroom. The ripple coefficient isn't just a performance metric; it's a proxy for the entire engineering trade space around processor selection, thermal management, and hardware cost.

The comparison to value iteration is instructive here. Classical approximate dynamic programming using value iteration requires discretizing the continuous state space into a grid, then iteratively updating the value of each grid point until convergence. For a two-dimensional state space with moderate resolution—say, 100 points per dimension—you have 10,000 states to evaluate. For each state, you must consider all possible actions (8 voltage vectors) and update the value based on the Bellman equation. With a discount factor close to 1, convergence can require thousands of iterations. The result is a lookup table that must be interpolated during online operation, adding latency and approximation error.

The LMI approach avoids both problems. By restricting the function class to quadratics, it implicitly extrapolates to all states—including those never visited during the offline search. There's no grid, no interpolation, no dimensional curse. And because the function is quadratic, evaluating it at an arbitrary state requires only the stored coefficients, not a search through a lookup table.

What's Next

The paper identifies several avenues for future work. The most immediate is experimental validation: all results currently come from simulation, and real motors introduce complications the model doesn't capture. Nonlinearities from magnetic saturation, inverter dead-time, measurement noise, and parameter uncertainty could all degrade performance relative to the simulated case. The researchers acknowledge this and note that experimental verification is planned.

There's also the question of how the method scales to more complex inverter topologies. The paper considers up to five voltage levels, but multilevel inverters with seven, nine, or more levels are used in high-power applications such as traction drives and renewable energy inverters. The number of voltage vectors grows combinatorially with the number of levels, which could challenge the LMI formulation. The researchers suggest that symmetry in the voltage vector set might be exploited to reduce the number of constraints, but this remains to be shown.

Another open question concerns the choice of objective function in the offline synthesis. The paper uses the spatial average of the value function over a ball of radius ε (Equation 26), which ensures the approximation is tight near the origin. Alternative objectives—for example, emphasizing performance near the typical operating trajectory or weighting states by their likelihood under the stationary distribution—could improve practical performance without changing the computational structure.

The conservativeness introduced by the S-procedure deserves more attention. Proposition 1 gives sufficient conditions for the Bellman inequality to hold, meaning some valid value function approximations might be excluded by the LMI constraints. The paper doesn't quantify how much conservativeness this introduces, and doing so would help engineers understand when the method is likely to perform close to the theoretical optimum.

Finally, there's the matter of adaptive operation. The current formulation assumes the motor parameters—resistance, inductance, magnet flux—are constant and known. In practice, resistance varies with temperature, and inductance varies with current due to magnetic saturation. An adaptive version that updates the value function approximation online (or switches between pre-computed approximations for different operating points) could extend the approach to more demanding applications.

For now, the paper establishes a compelling proof of concept. It shows that the infinite-horizon optimal control of a three-phase motor can be approximated efficiently using convex optimization, that the approximation is tight enough to give near-optimal performance, and that the resulting controller is simple enough to implement on modest hardware. Whether this approach finds its way into production electric vehicles will depend on experimental validation, industrial adoption, and the continued pressure to improve efficiency in electrified transportation. But the mathematics is sound, the results are promising, and the idea that you can compute optimal control policies offline and apply them online with minimal effort is a conceptual advance that extends well beyond motor control. The same techniques could, in principle, be applied to any switched system where the computational cost of FCS-MPC has been prohibitive—grid-tied inverters, multilevel converters, hybrid electric vehicle powertrains, and beyond.

The future of electric motion may ultimately depend less on faster processors than on smarter algorithms. This paper is a step in that direction.