Cross-SNR Dynamic DC-Bias Prediction in DCO-OFDM Using a Domain-Informed Mixture Density Network

Light is becoming a new way to carry information. Just as radio waves transmit data to our phones, the light from a standard LED bulb can be flickered so quickly that the human eye cannot see it, yet a receiver can decode the pattern into digital information. This technology, often called Li-Fi, offers a way to communicate without using radio frequencies, keeping signals contained within a room and avoiding interference. To make this work efficiently, engineers use a method where the light intensity changes to represent complex data patterns. However, because light bulbs can only get brighter or stay dark—they cannot go "negative"—the signal must be shifted upward to ensure it never dips below zero. This shift is achieved by adding a constant amount of electrical power, known as a direct current bias.

The size of this shift is a delicate balancing act. If the shift is too small, the bottom of the signal gets chopped off, creating errors in the data. If the shift is too large, the system wastes energy and demands more power from the hardware than necessary. The challenge is that the perfect amount of shift changes constantly depending on the quality of the connection and the complexity of the data being sent. For years, engineers have relied on fixed settings or complex calculations to guess the right amount, but these methods often fail when conditions change unexpectedly. A team of researchers has now developed a new approach that learns to predict the ideal shift in real-time, adapting to the environment with a level of precision that fixed rules cannot match.

In a study published in October 2026, researchers Muhammad Bilal Rehman, Momina Mir, and their colleagues at the National University of Modern Languages and the Institute of Space Technology in Pakistan tackled this problem using a form of artificial intelligence. They focused on a system called DCO-OFDM, which is a standard way of organizing light signals for high-speed communication. The core difficulty they addressed was that the optimal power shift depends on two things: the specific shape of the light signal at that moment and the current quality of the connection, measured as the signal-to-noise ratio. Traditional methods often try to find a single best number for a given situation, but the researchers realized that the relationship between the signal, the noise, and the required power shift is too complex for a simple guess.

To solve this, the team created a new type of computer model called MDHyperNet. Instead of trying to predict one single number for the power shift, this model predicts a range of likely values, understanding that there is a degree of uncertainty in every connection. The model was trained on a dataset where the researchers simulated thousands of different scenarios. For each scenario, they calculated the perfect power shift by weighing three competing goals: keeping data errors low, minimizing the peak power required, and reducing the noise caused by chopping off parts of the signal. They used a mathematical method called TOPSIS to find the best compromise among these goals, creating a set of "gold standard" answers for the model to learn from.

The most significant test of this new system was its ability to handle situations it had never seen before. In many machine learning studies, a model is tested on data that looks very similar to what it learned from. Here, the researchers used a strict testing method where they completely removed entire ranges of connection quality from the training data. They trained the model on some signal qualities and then asked it to predict the correct power shift for signal qualities it had never encountered. This is a difficult task, similar to asking a driver who has only practiced on dry roads to navigate a sudden, heavy snowstorm without prior experience.

The results showed that the new model succeeded where older methods failed. When tested on these unseen conditions, the traditional methods made large errors, often guessing a power shift that was far too high or far too low. In the most difficult test cases, where the connection quality was at the extreme edge of what was known, the old methods were off by as much as fifty percent. In contrast, the new model stayed remarkably accurate, with errors staying below two percent. The researchers found that the key to this success was not just the complexity of the model, but the fact that they built a specific rule into its design: they told the model explicitly that the power shift depends on the connection quality. This "domain-informed" approach allowed the model to understand the physics of the situation rather than just memorizing patterns.

To prove that this accuracy mattered in the real world, the researchers ran full simulations of a complete communication system. They compared their adaptive system against standard setups that use a fixed power shift of 7, 10, or 13 units. The fixed systems always struggled; a setting that worked well for a strong connection caused waste and errors when the connection weakened, and vice versa. The new adaptive system, however, adjusted the power shift for every single data symbol. This allowed it to maintain a low error rate across the board while using less power and generating less noise than any of the fixed settings. The system effectively found a middle ground that no single fixed setting could achieve.

The study also revealed which parts of the new model were most important. The researchers tested versions of the model with and without the specific rule about connection quality, and with different numbers of possible outcomes. They discovered that the rule linking the power shift to the connection quality was the primary reason the model could generalize to new situations. The part of the model that predicted a range of values was less critical for raw accuracy but helped the system better match the distribution of the ideal answers. This suggests that for this specific problem, understanding the physical relationship between variables is more important than simply having a more complex neural network.

While the results are promising, the researchers note that these findings come from computer simulations. The tests assumed a clean environment without the interference of sunlight or the non-linear behavior of real-world light bulbs, which are factors present in actual Li-Fi deployments. The model was also trained on data generated by a specific optimization method, meaning that if the priorities of the system changed, the model would need to be retrained. Despite these limitations, the work demonstrates a clear path forward. By combining a deep understanding of the physical constraints of light communication with a flexible learning model, the researchers have shown that it is possible to create a system that adapts dynamically to its environment. This approach moves beyond static settings, offering a way to keep optical wireless links efficient and reliable even as conditions change.