Buildings are responsible for an enormous share of global electricity demand, and the way their air conditioners, lights, and heating systems are scheduled can make the difference between a grid under strain and one running smoothly. For years, engineers have tried to optimize these schedules with evolutionary algorithms and reinforcement learning, but the problem becomes brutally hard when dozens of buildings share a single electrical network and must be coordinated simultaneously. A new study published in Discover Artificial Intelligence proposes a strikingly different approach: instead of searching step by step for the best schedule, a generative artificial intelligence model learns to paint one directly out of pure noise, guided by a map of how the buildings are physically wired together.

The framework, called DMCBO, short for Diffusion-driven Multi-objective Collaborative Optimization for Building Clusters, was developed by Yalan Zheng of Zhengzhou Technical College and Taixiang Yin of North China University of Water Resources and Electric Power. Their central insight is that the joint scheduling of equipment settings and load distribution across an entire cluster of buildings can be treated as a high-dimensional data sample, much like an image. Diffusion models, the same family of generative techniques behind modern text-to-image systems, work by progressively corrupting data with Gaussian noise and then learning to reverse that corruption. The researchers reformulated multi-objective building optimization as exactly this kind of conditional denoising process: starting from random noise, the model strips away noise step by step until a complete, Pareto-optimal collaborative schedule for every piece of equipment in every building emerges.

Technically, the forward diffusion process is a fixed Markov chain that adds tiny amounts of Gaussian noise to real scheduling data over one hundred steps, with the noise coefficient increasing linearly from 0.0001 to 0.02 until the original data becomes indistinguishable from standard Gaussian noise. A neural network is then trained to predict the noise that was added at each step, given the noisy schedule and a conditional encoding vector describing the current situation. Once trained, the model can generate fresh solutions at inference time by sampling pure noise and running the reverse process, refining it over one hundred denoising steps into a schedule tailored to the prevailing weather, occupancy, and grid conditions. Because this generation is a gradient-based learning process rather than an iterative search, it avoids the local optima traps that plague traditional multi-objective evolutionary algorithms and the reward oscillations that destabilize deep reinforcement learning.

What makes the system genuinely cluster-aware is its use of graph neural networks. The researchers model the building cluster as a graph mapped onto a radial distribution microgrid topology, where nodes represent buildings and edges encode physical constraints such as maximum line thermal capacities and transformer limits. A two-layer graph convolutional network passes messages between neighboring buildings, fusing spatial-topological dependencies and network constraints into a single graph embedding. This embedding is concatenated with an adjustable preference weight vector to form the condition that steers the denoising trajectory. In effect, when the model generates a control strategy for one building, it simultaneously perceives the state of its neighbors and the limits of the shared electrical infrastructure, suppressing systemic coordination errors from 0.18 down to 0.06 in the team’s experiments.

The multi-objective aspect is handled through an elegant scalarization trick. Each objective, such as energy consumption, operating cost, or thermal comfort, is normalized to a dimensionless value between zero and one using Min-Max normalization based on historical operational bounds. The normalized objectives are then blended into a single scalar condition via a weight vector whose components sum to one. During training, these weights are sampled from a uniform Dirichlet distribution, exposing the model to the entire space of possible preferences so it learns the distribution of the full Pareto front. At deployment, a decision maker can slide the weight vector like an adjustment knob, instantly generating schedules that emphasize energy saving, comfort, or a balanced compromise, all within the same trained model.

The denoising backbone itself is a U-Net adapted to process one- and two-dimensional building scheduling tensors. Diffusion time steps are embedded through sinusoidal positional encodings, while the graph-derived condition is injected into every residual block via Adaptive Layer Normalization. A bottleneck layer with multi-head self-attention captures long-range dependencies across the entire daily scheduling horizon, and symmetric skip connections preserve high-resolution information during reconstruction. The authors trained the model for 300 epochs on dual NVIDIA Tesla V100 GPUs, and the loss curves dropped from 0.75 to below 0.03 with training and validation curves closely coinciding, indicating no overfitting. As the reverse diffusion step count decreased from 90 to 10 during analysis, the noise-prediction error distribution became sharply concentrated at zero, demonstrating extremely high denoising precision at the end of generation.

Validation was carried out on the open-access Building Data Genome Project 2 dataset, which supplied real hourly weather records, operational schedules, and annual electricity, cooling, and heating load profiles across diverse climate zones. Because that dataset provides metered energy data rather than physical building models, the team used United States Department of Energy Commercial Reference Buildings as templates and calibrated them in EnergyPlus, a high-fidelity thermodynamic simulation platform, to create a digital twin consistent with the real data. All results were benchmarked against a rule-based control strategy with static thermostat setpoints of 25 degrees Celsius in summer and 22 degrees in winter, and every experiment was repeated ten independent times to confirm statistical significance.

The headline numbers are impressive. In high-demand summer conditions, DMCBO achieved an average energy-saving rate of 28.6 percent and cut peak load by 38.5 percent, shaving the afternoon peak from roughly 560 kilowatts down to about 440 kilowatts and saving approximately 142.80 US dollars per day. Cumulative daily power savings exceeded 1,000 kilowatt-hours, compared with 400 kilowatt-hours for a traditional collaborative strategy and just 120 kilowatt-hours without coordination. In winter, peak shaving and valley filling reached about 39 and 42 percent respectively, far above the 22 and 25 percent of conventional methods. When the solution sets were projected into objective space, DMCBO dominated the baselines: it reached a hypervolume indicator of 0.95 within 12,000 function evaluations, versus 0.85 for NSGA-II and 0.70 for MOPSO, while keeping operating costs between 140 and 160 dollars against more than 210 dollars for deep reinforcement learning and 260 to 300 dollars for a GAN-based generative approach.

Robustness and scalability proved equally compelling. Under injected stochastic disturbances mimicking sudden weather changes and occupancy swings, the framework confined power fluctuations to a narrow band of plus or minus 5 kilowatts, whereas the deep reinforcement learning baseline swung between minus 15 and plus 20 kilowatts. Scaling from 5 to 50 buildings, the global energy-saving rate actually rose from 28.5 to 31.2 percent, because larger clusters unlock more load-complementary opportunities that the graph encoder can safely exploit within encoded grid boundaries. Computation time grew to 9.8 seconds at the 50-building scale, but that remains comfortably within the 15-minute control horizon used by real building management systems, and the hypervolume index slipped only slightly to 0.88. Ablation experiments confirmed that both the graph neural network module and the choice of 100 diffusion steps, which yields a 50-millisecond inference latency, were essential to this balance of speed, accuracy, and stability.

The authors are candid about limitations. The framework has so far been validated only in simulation, and EnergyPlus simplifies sub-minute mechanical actuation latencies that real HVAC hardware would introduce; they propose cascading PID loops and predictive safety margins to absorb those delays in practice. Data quality, generalization to residential buildings and extreme climates, the computational cost of scaling to city-sized networks, and the black-box nature of the generative process all remain open challenges, with explainable AI and transfer learning flagged as future work. Still, the architecture is designed to slot into standard building management systems as a supervisory layer, transmitting setpoints over protocols such as BACnet or Modbus. If the simulation gains survive contact with physical buildings, the idea of generating optimal energy schedules from noise, guided by a graph of the grid itself, could become a powerful new tool in the push toward decarbonized cities.

Subject of Research: Conditional diffusion models with graph neural networks for multi-objective energy optimization in building clusters

Article Title: Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks

Article References: Zheng, Y., & Yin, T. (2026). Diffusion model-driven multi-objective collaborative optimization for building energy management using graph neural networks. Discover Artificial Intelligence, 6(1), Article 1383. https://doi.org/10.1007/s44163-026-02140-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02140-z

Keywords: diffusion models, graph neural networks, building energy management, multi-objective optimization, EnergyPlus, smart buildings, Pareto optimization, deep learning, energy efficiency, HVAC control, microgrids, generative AI

One of the most popular explanations of human behavior has been that somewhere beneath our rational mind lurks an ancient “lizard brain”—a primitive evolutionary holdover responsible for instinct, emotion, fear, and aggression.

There is just one problem. Neuroscientists have long known that the lizard brain story is wrong. Humans did not evolve a sophisticated new brain by simply stacking it on top of an older reptilian one.

But a new study suggests the enduring lizard brain myth may have inadvertently pointed to something more interesting. Mammalian brains do appear to be organized around two broad computational systems—but not an ancient reptilian brain located beneath a newer, more sophisticated one. Instead, the divide appears to reflect two fundamentally different ways of organizing and processing information.

Published in Science Advances, the research combines anatomical measurements from 182 mammalian species with experiments using artificial neural networks. The researchers found a striking evolutionary trade-off between the neocortex and a collection of structures associated with smell, memory, and the limbic system.

Instead of representing “new” and “old” brains, the researchers say these regions may have evolved different wiring styles optimized for solving different kinds of computational problems.

“Rather than an ‘old’ brain and a ‘new’ brain, our results suggest two systems wired for different kinds of computation,” lead author Dr. Nabil Imam, a researcher at the Georgia Institute of Technology, told The Debrief. “The neocortex is wired to preserve spatial relationships—for example, neighboring points in the visual world are analyzed by neighboring parts of the cortex. The limbic system is more diffusely connected and supports holistic associations across different pieces of information.”

“So, the division between the neocortex and limbic system is about how the two systems are wired,” Dr. Imam added.

This distinction offers a very different picture from the famous lizard brain story.

The lizard brain concept grew from the so-called triune brain theory, which portrays the human brain as having evolved in distinct layers. A primitive “lizard brain” handling basic instincts, a mammalian limbic system producing emotion, and, finally, a newer neocortex responsible for higher thought.

The problem is that neuroscience has long since overturned that evolutionary story. Vertebrate brains did not evolve by progressively stacking newer brain regions on top of older, more primitive ones. Instead, vertebrates share essentially the same broad organizational plan, with existing regions changing in size, connectivity, and function over evolutionary time.

The new study suggests, however, that the mammalian brain may be divided very differently—not by evolutionary age, but by how different neural systems are wired to process information.

Across 182 species from 10 taxonomic groups, animals with an unusually large olfactory system tended to have a smaller-than-expected neocortex. Animals with an expanded neocortex tended to show the opposite pattern.

The nine-banded armadillo provides a particularly dramatic example. Armadillos are largely nocturnal and depend heavily on smell while foraging. As researchers note, the armadillo’s olfactory system was approximately 1.5 times larger than expected for its overall brain size. Its hippocampus—a region central to memory and spatial processing—was about twice its expected size.

Its neocortex, meanwhile, was reduced by a factor of 2.8. Primates such as squirrel monkeys sit closer to the other end of the spectrum. They rely heavily on vision and devote a large portion of their expanded neocortex to processing visual information.

At first glance, that could look like evolution trading smell for sight. But the researchers noticed something more intriguing. The hippocampus grew and shrank alongside the olfactory system, even though memory is obviously not another form of smell. Other structures traditionally grouped into the limbic system, including the amygdala and septum, also tended to change together.

That suggested the important difference might not be what these regions process, but how they are wired to process it.

Two Very Different Ways to Build a Brain

Vision offers a relatively straightforward example. Neighboring points in the visual world are generally processed through neighboring areas of the visual system. Touch operates similarly. Locations beside one another on your body retain spatial relationships as the nervous system processes them. Hearing also uses orderly mappings, including those based on sound frequency.

The result is what neuroscientists call “spatiotopic” organization. In simple terms, the brain preserves relationships between incoming information by arranging it into orderly maps.

However, smell works differently. Odors are represented through combinations of activity distributed across networks rather than being laid out in the same neat spatial fashion.

The researchers suspected that this difference might help explain the anatomical trade-off they had identified across mammals. To test the idea, they turned to artificial neural networks trained to handle vision, touch, hearing, and smell.

They then mapped the networks’ internal models onto artificial two-dimensional surfaces resembling sheets of cortex.

Results showed that networks optimized for vision, touch, and hearing produced smooth, highly ordered maps. On a scale in which 1 represented strong spatial order, the visual network scored 0.90, touch scored 0.86, and hearing scored 0.89.

Yet, the smell network scored just 0.11. Rather than producing a smooth map, information was scattered across a fractured, distributed representation. The researchers even found that imposing localized connectivity on the olfactory model harmed its ability to perform its task.

Researchers then examined memory. Using an artificial network designed to reproduce some relational computations associated with the hippocampus and entorhinal cortex, researchers again saw a fractured, distributed representation resembling the olfactory system rather than vision, hearing or touch.

For Dr. Imam, one of the most surprising results was that artificial networks arrived at an organization resembling real brains despite learning in profoundly different ways.

“Artificial networks are trained by an algorithm that propagates errors through the network, whereas brains are shaped by evolution and through localized changes in synapses,” Imam explained. “The fact that both nevertheless arrive at similar organization suggests that the structure of the information itself—whether visual, auditory, tactile, or olfactory—determines what kind of wiring works best.”

No “Lizard Brain,” But Evolution May Be Trading One Kind of Brainpower for Another

Researchers next wanted to know whether competition between these two wiring styles could reproduce the anatomical patterns seen across real mammals.

They created an evolutionary simulation containing five artificial networks. Vision, hearing, and touch occupied one computational domain based on localized, spatial organization. Smell and hippocampal memory occupied another based on distributed connectivity.

Then an evolutionary algorithm redistributed computational resources between them.

When selection favored vision, resources moved toward the spatially organized networks. When selection favored smell, the distributed networks expanded.

But something interesting occurred to the simulated hippocampal network. Even though researchers weren’t directly selecting for memory, the hippocampal network grew along with the olfactory network because both shared the same broad computational architecture. As resources shifted toward the other system, both contracted.

That closely resembled the pattern found across real mammalian brains.

“It suggests that evolution may not be selecting individual brain regions one at a time, but rather the functions produced by whole patterns of wiring,” Dr. Imam explained. “Because regions that share similar wiring tend to grow or shrink together, selection for one kind of function comes at the expense of another, given that the system must operate with limited energy and space.”

That does not mean an animal with a larger neocortex is simply more intelligent than one with a larger limbic system. Nor does the distributed system represent a primitive emotional brain fighting against a rational one.

Instead, the two architectures appear suited to different computational problems. Evolution may have repeatedly adjusted how much neural real estate mammals devote to each.

Nevertheless, researchers caution that their models are intentionally simplified. Anatomical measurements came from several datasets; some taxonomic groups are better represented than others, and artificial neural networks are not replicas of biological brains.

The evolutionary model is intended to demonstrate that a relatively simple developmental mechanism could produce the observed trade-off, not to reconstruct exactly how mammalian brains evolved.

Still, the research raises another interesting possibility extending beyond the demise of the “lizard brain.” Our brains may arrive in the world with considerably more computational structure already built in than the familiar idea of a blank slate implies.

“The brain may be much less of a blank slate than we often imagine,” Dr. Imam told The Debrief. “Evolution and development build useful assumptions into its wiring before experience-driven learning.”

It is also another example of something scientists are increasingly encountering across very different areas of biology: evolution can produce remarkably sophisticated structures and systems that human engineers are just starting to understand, much less reproduce.

In a recent study covered by The Debrief, researchers discovered that cable bacteria construct microscopic electrical pathways from nickel-rich nanoribbons arranged remarkably like braided electrical wires. The naturally grown structures may achieve extraordinary conductivity while remaining flexible, properties that researchers are now interested in replicating for bio-based electronics.

The mammalian brain represents an enormously more complicated example, but the underlying lesson may be similar. Evolution does not begin with a blank slate each generation. It can build useful solutions into biological architecture itself, leaving subsequent learning along with adaptation to operate within structures refined over immense spans of evolutionary time.

That could also help explain one of the most striking differences between biological intelligence and today’s artificial intelligence. Modern AI systems can require enormous datasets to acquire abilities that biological brains learn from comparatively limited experience.

“That built-in structure may explain why brains can learn efficiently from relatively little experience, whereas most artificial networks require enormous datasets to learn useful functions,” Imam said. “The idea has deep roots in philosophy of mind, but only recently have neuroscience and computer science begun to provide concrete demonstrations of how such built-in structure can aid function.”

The recent study, “Dual computational systems in the development and evolution of mammalian brains,” appeared in Science Advances.

Tim McMillan is a retired law enforcement executive, investigative reporter and co-founder of The Debrief. His writing typically focuses on defense, national security, the Intelligence Community and topics related to psychology. You can follow Tim on Twitter: @LtTimMcMillan.  Tim can be reached by email: tim@thedebrief.org or through encrypted email: LtTimMcMillan@protonmail.com