16
Starry Sky Aerospace Simulation Center.
Soon, the three-module joint simulation environment was set up.
It was not just a single server room, but a closed area spanning half a floor.
Inside, a cluster was formed by more than twenty high-performance computing nodes.
Here, a complete multi-physics field coupled simulation platform rewritten from the underlying layer—covering orbital dynamics, fluid heat transfer, power networks, and structural mechanics—could be run.
Moreover, the biggest difference from the old system used previously was that this time the benchmark data source was no longer ground experimental data.
Instead, it was the real telemetry records of the experimental module in orbit for more than 180 days.
Zhou Heng, the head of the simulation group, stood at the entrance of the server room, pointing at a row of blinking cabinets, and said,
"The two-module combination is currently running in space. We just need to feed its real-time telemetry in every day to keep the simulation environment synchronized with the real space."
"It is equivalent to building a digital twin on the ground identical to the one in space."
Hearing this, Jiang Huaixing nodded and then asked,
"What is the synchronization accuracy?"
"The deviation of key parameters is controlled within five-thousandths. But the problem is—"
Zhou Heng's expression was not very relaxed.
"However, there are always variables in real space that we do not know."
"For example, device aging caused by space radiation, cumulative impacts of micron-level debris, material outgassing effects, and so on."
"We cannot observe these problems, but the equipment in space can feel them."
"Therefore, the simulation environment can only simulate what we already know."
"You mean what we simulate is always lagging behind."
"Right."
"But this is already much better than before."
"At that time, we didn't even know about an offset of three-thousandths; now at least we know which direction to supplement data in."
...
The first three-module full-link joint simulation ran for seventy-two hours.
The AI took over global scheduling, and the power, thermal control, and attitude control of all three modules were handed over to it for allocation.
However, the final result was not very good.
"The success rate is sixty-one percent."
The AI Group Leader looked at the simulation report on the screen, brows knitted together, and said,
"It is even lower than the two-module setup."
"And it's not just a little lower; it's lower by thirty-six percentage points."
"Where does the problem lie?"
"It's not power beating."
"Because beat oscillation is well controlled under the upgraded algorithm."
"We have compressed the convergence time to 0.4 seconds."
"The problem should be right here."
Afterwards, he brought up a heat map.
The resource scheduling frequency of the three modules in the simulation was represented by color shades.
Among them, dark red represented an extremely high decision-making frequency of the AI on that module.
"After the three modules are connected to the grid, the dimension of resource scheduling changes from two-dimensional to three-dimensional."
"The AI needs to make decisions on three things simultaneously: power distribution, thermal control scheduling, and attitude control."
"And these three things are mutually coupled."
"For example, adjusting the solar panel angle will affect power generation, but the solar panel angle also affects the module's thermal balance."
"However, changes in thermal balance will in turn affect equipment power consumption, and equipment power consumption will conversely affect power demand."
"Circular coupling."
"Right."
"In the two-module scenario, the AI only needed to make decisions in a two-dimensional plane, and the path was very clear."
"But in the three-module scenario, it is equivalent to turning into a three-dimensional spatial volume, and the number of decision-making paths grows exponentially."
"So it starts to 'hesitate'."
"Then it repeatedly jumps back and forth between two similar scheduling schemes, causing the system state to oscillate continuously."
"Why would this happen?"
"Because its training objective function is still 'maximize system stability' written from the two-module era."
"In the three-module scenario, this objective function itself has gray areas."
"For example: power priority or thermal control priority?"
"When the two conflict, the AI does not have a clear basis for trade-offs."
Jiang Huaixing thought for a moment and asked again,
"Then why didn't they conflict during the two-module period?"
"Because there was a large surplus during the two-module period."
"There was a surplus in power and a surplus in thermal control; the AI didn't need to make trade-offs, and all indicators could be met simultaneously."
"And the three-module T-shaped configuration wiped out the surplus, forcing the AI to make a choice."
"Then it got dazed."
"So does that mean it's not that our algorithm is no good, but that the training objective is insufficient."
"Right."
"In the two-module era, 'maximizing stability' was enough."
"Whereas in the three-module era, a set of decision-making logic with priorities is required."
"For example,'systems related to life safety have absolute priority, scientific payloads come second, and comfort functions come last'."
"Then what if we add priorities?"
"It has already been added."
The AI Group Leader brought up the second figure.
"After adding it, the success rate increased from sixty-one percent to seventy-eight percent."
"But it's still not enough."
"What do you mean?"
"Uh, for example, a certain scientific payload is doing a critical experiment when suddenly power becomes tight."
"A hard priority will directly cut off the payload to protect the core system."
"But in fact, if we only reduce the frequency by twenty percent, it can pull through, and the experiment can continue."
"Between this hard cut and flexible degradation, the AI cannot judge well yet."
"Then let it learn flexible degradation first?"
"That is the problem."
The AI Group Leader's expression looked a bit tired.
"Flexible degradation requires judging 'how much to reduce' and 'when to restore'."
"And this requires predicting the system state changes in the next few minutes."
"And our simulation environment—"
Speaking of this, he paused, then pointed to the server room and said,
"has limited predictive capability."
"The disturbances in real space are continuous and random, and our simulation can only give discrete predictions based on known variables."
"Therefore, the flexible degradation strategy learned by the AI in the simulation may not be so applicable when it gets to space."
Silence fell for a moment.
Then Jiang Huaixing asked again,
"What about the on-orbit data?"
"The two-module combination has been running for more than a month now."
"We have it."
"We have already fed all the real scheduling logs of the two-module setup for these forty-plus days."
"However, the decisions made by the AI in the two-module era were all 'optimal choices when there was a surplus'."
"It has never experienced truly resource-constrained scenarios."
"This is equivalent to telling someone who has never been poor how to save money; they won't understand."
"Then let it be poor once."
"In the simulation."
"We artificially inject resource-constrained working conditions."
"Cut the solar panel area by twenty percent, increase the radiator shielding ratio to forty percent, and add thirty percent to the load power."
"In this way, we should be able to create an extreme environment that is 'poor to the extreme'."
"Then let the AI struggle repeatedly inside until it learns flexible degradation."
"But the success rate will look very ugly this way."
"Failing a thousand times in the simulation can mean one fewer failure in space."
"The success rate looks ugly? Ugly is ugly."
"The data doesn't look good, but the account has to be settled in space."
"It's not that we aren't afraid of difficulties, but we cannot skip this step."
"If it doesn't pass a single run, then don't put it on the real machine yet."
"Being a little slower is fine; an accident is the big deal."
"Make sure to record today's data clearly first, and whoever is in charge of the module takes responsibility for it."
...
After the extreme working conditions were injected, the simulation ran for three hundred rounds.
In the first fifty rounds, the AI was almost completely wiped out—either power collapsed, thermal control became unbalanced, or attitude became unstable.
But by the one-hundredth round, a turn for the better began to appear.
...
