Having spent the night analyzing 20 gb of samples, I discovered that the publicly leaked data contains far more than the cybersecurity researchers who originally studied it had realized.
Today we will examine in detail the simulations of direct target engagement by various munitions found in these files. On this topic, the leak contains two large PDF files (detailed reports on the simulation work carried out by Chinese physicists) and a couple of animations, purely as a teaser for what is being offered for sale. The actual simulation files, detailed blueprints, models, etc. are not present in the conditionally publicly available leak samples.
Based on the information presented, the Chinese have had the good sense to break free from their ANSYS dependency, and are already actively using their homegrown application package GalaxyEDS for the majority of their calculations - and with a richer, more up-to-date library of equations of state, materials, and so on. Additionally, it has built-in support for hierarchical/modular simulation models for large-scale pipeline calculations of thousands of different cases - something that the ancient Autodyn lacks, since it was originally designed for a somewhat different purpose.
Autodyn is used here purely for verification of GalaxyEDS's performance, to prove that the homegrown solution produces results with equivalent accuracy. Judging by the screenshot, the .ad files shown are precisely the verification files:
In addition to the raw simulation tools, they have also built a pipeline for cleaning, processing, and analyzing all the data obtained through these simulations.
Ground targets
For simulating the engagement of ground targets, the following munition was modeled, based on the likeness of a certain real rocket:
However, no air-to-ground munitions, artillery shells, or MLRS rockets of a similar caliber - let alone ones with tungsten fragments - are anywhere close to being in PLA service based on public data. So one can assume that these preliminary ground target engagement simulations either use a model of an experimental munition whose very existence has not yet been made public, or this is a simulation of ground vehicle engagement by an Israeli LAR-160 (correct me if you are familiar with the LAR-160 warhead and it doesn't look like this).
The target chosen was a HIMARS, which clearly has the Chinese worried - a logical choice in the context of examining counter-battery effectiveness:
They went all out with this model - every more-or-less significant component is simulated with carefully selected material properties, down to the steering wheel, air filters, various onboard computers, and the plastic seats in the cab. This is clearly groundwork for numerous simulations aimed at determining the probability of disabling individual target subsystems:
The system should be able to perform response analysis on different levels of structures. Especially for critical functional components (such as fuel tanks, communication bays, and power systems), it should ensure that damage assessments have realistic physical meaning and guiding value. Finally, in terms of damage assessment, it needs to have a comprehensive assessment capability for structural damage and functional failure, capable of outputting damage parameters including indicators such as structural penetration and localized fracture, while simultaneously determining whether the target has lost its combat functionality, providing a reliable basis for judging the effectiveness of the attack, assessing battle damage, and adjusting subsequent plans.
In this particular simulation, special attention was paid to the probability of disabling the FCS and the aiming mechanism, as systems critical to the operation of this MLRS.
Naval surface targets
For simulating the engagement of American aircraft carriers, a "certain missile" was modeled, which is obviously the quasi-ballistic YJ-21:
Among previously non-public data, the shape and dimensions of the kinetic penetrator made of tungsten alloy are clearly visible:
The target chosen was an American Nimitz-class aircraft carrier:
Since we are now dealing with entirely different scales and energies, modeling individual plastic seats is no longer on the table. The superstructure is modeled fairly roughly, because even with their supercomputer, the Chinese simply don't have enough computing power for highly detailed calculations:
The only thing that raises an eyebrow is the Su-47 model on the deck:
Apparently, the researchers were not tasked with examining in detail the damage to deck aircraft from anti-ship missile fragment impacts, so the first available model that roughly corresponded to the volume of modern aircraft was used as a placeholder - purely for demonstrating blast wave propagation near the carrier's superstructure in a scenario where an anti-ship missile strikes the island directly (the air domain was simulated only in that area). I believe that by the time the pipeline for running hundreds of simulations is launched, the aircraft model will be replaced with a more detailed model of an F-35C.
What is interesting is that the author does not envision the need to simulate an anti-ship missile hitting "the carrier" in general, but rather assumes that this hypersonic sledgehammer will be able to select a point of impact with accuracy down to half a meter:
To effectively destroy the command and communication equipment on the island, the initial strike position during penetration is located on the outer perimeter of the rear half of the fifth layer of the island superstructure. The warhead's impact velocity is approximately Mach 5.9 (2000 m/s), meeting the performance indicator of "simulation calculations based on GalaxyEDS covering ultra-high-speed impact penetration."
I am not an expert in guidance systems, so I welcome your comments - is this even possible?
The analysis of this simulation devotes significant attention not merely to the question of "how thoroughly the superstructure was destroyed," but also to issues of blast wave propagation and reflection, pressure distribution at different stages of the explosion, etc. - which could potentially help them capture more subtle effects of warhead-on-target interaction than simply "strap on the biggest firecracker and send it."
The result of this simulation is, nevertheless, obvious - the superstructure is destroyed, the carrier is knocked out of action, most likely sunk. However, one should not laugh at the obviousness of the answer - precise simulations of munition-target interaction help miniaturize the munition, and if the existing munition is clearly overkill, such simulations help find the optimal parameters that will be just sufficient.
I have no doubt that for the final tests of the munition, the Chinese can afford to build a full-scale replica of the superstructure at 1:1 scale, but in order to find the size at which the munition should be built - simulations are indispensable.
Underground targets
For simulating the engagement of a concrete bunker, the GBU-24 was chosen:
Most likely, the munition was chosen based on its similarity to comparable Chinese bunker-buster bombs built in a similar form factor.
Three different types of bunkers were modeled as targets for different mission profiles:
The documents also mention simulations of nuclear weapons effects on similar fortifications, but only as one of the possible avenues for the project's further development.
There is not much to comment on here, but there are two cool animations - here is one of them:
Project prospects and timelines
All of these simulations are being conducted as part of a single project with highly ambitious goals. The main objective of the project is to create a pipeline for running thousands of simulations with different munitions, targets, and engagement parameters:
The land mobile vehicle target simulation model is a comprehensive structural system. In damage simulation analysis, the mobile vehicle can be modeled in a modular, hierarchical manner: the first level is the component level, including six damage scenarios such as armor plates, tires, fuel tanks, missiles, support frames, and the driver's cab; the second level is the assembly level, including four damage scenarios such as protective components, chassis components, power components, and damage element components; the third level is the whole vehicle level, evaluating one whole vehicle level damage scenario. The land mobile vehicle target simulation model has a total of 11 damage scenarios. At the same time, four different speed conditions, four different angle conditions, and four different position conditions are planned for each damage scenario. Therefore, the damage simulation calculation conditions of the land mobile vehicle target simulation model are: (6+4+1)*4*4*4=704.
This is needed to produce a dataset that can be used both for developing more effective and miniaturized munitions, and for training a combat AI that would now have sufficiently precise information about the probability of damaging / destroying various targets across all possible scenarios:
Effective damage data for each state in the spatial dimension enables in-depth analysis of damage mechanisms, effect assessment, and strike plans. By setting different initial conditions (such as impact velocity, impact angle, ambient temperature, etc.) and boundary conditions through numerical simulation, hundreds or even thousands of different damage conditions are systematically simulated (one calculation under defined boundary conditions and simulation model constitutes one condition). First, a simulation model is established to accurately reflect the geometric characteristics and material constitutive relations of the target. Then, using a high-performance computing platform and employing explicit dynamic algorithms, high-fidelity simulations of physical processes such as explosion impact, projectile penetration, and structural failure are performed, generating simulation results (including spatiotemporal evolution data of multi-physics fields such as stress field, strain field, displacement field, and temperature field). This allows for the accumulation of terabytes of simulation data in a short period, covering multiple damage scenarios while avoiding noise interference from experimental measurements. This yields high-quality, large-scale physical simulation data, providing massive data support for intelligent model training.
Here are more quotes if you doubt that "intelligent model training" refers specifically to AI:
Numerical noise (such as high-frequency oscillations and numerical dissipation errors) and non-physical components (such as outliers that violate conservation laws) can directly contaminate simulation results. Deep learning models are highly sensitive to the quality of input data. Data noise may cause algorithms to misinterpret noise patterns as physical laws (such as mistaking high-frequency oscillations for turbulent fluctuations), leading to incorrect predictions. Therefore, it is necessary to detect oscillations and process non-physical noise in simulation data.
Primarily used for intelligent model training. High-quality datasets are the foundation and key to model training. Training intelligent models based on accurate numerical simulation results enables them to understand physics, achieving a leap from traditional physical simulation to data-driven intelligent analysis, and supporting rapid analysis and decision-making regarding battlefield damage.
The December 2025 date mentioned in the documents marks the beginning of this large-scale simulation project - essentially, what we see in the samples is only a proof of concept. Conducting the necessary thousands of simulations and compiling them into convenient datasets is planned only for May 2027:
(1) Milestones in two stages
The project plans to establish two milestone phases, in 2026 and 2027, with the specific phases outlined below:
Phase 1 (December 31, 2026): Complete a cumulative total of 1000 target destructions.
Data from simulation of damage under working conditions.
Phase Two (May 31, 2027): Achieve a cumulative total of 1500 target damages.
Simulation results of damage under working conditions.
The issue of some models used in these demonstrations not being entirely appropriate for the context - such as the Su-47 or GBU-24 - is more a problem of "we threw in whatever models we had at hand," and will be resolved in later stages of the project's development:
During the implementation of the project, firstly, there were difficulties in calculating the target damage scenario composed of damage elements and damage targets, such as the lack of model geometric structure data and material property data;
...
Countermeasures: Regarding technical risk one, this project aims to address the issues of missing model geometric structure data and material property data by reversing relevant data through model analysis, thus refining the internal geometric structure and material parameters of the simulation model.
That the project will indeed scale up rather than quietly die - is beyond doubt. Even this pilot stage of the project to build a pipeline for producing synthetic terminal ballistics data involves 17 people, a budget of $2 million and above, and is led by a heavyweight of Chinese political physics:
Yang Canqun is a researcher, doctoral supervisor, deputy chief designer of the Tianhe supercomputer project at the National University of Defense Technology, director of the National Supercomputing Center in Tianjin, deputy director of the CAE Committee of the Chinese Simulation Society, and a leading talent in military science and technology. He previously served as deputy director of the Computer Institute and director of the Software Institute at the National University of Defense Technology. He has long been engaged in high-performance computing technology research, undertaking key national projects such as the Galaxy and Tianhe supercomputers, and was responsible for the development of supercomputing software systems. He achieved independent control over supercomputing basic software development, made significant breakthroughs in core industrial software, and achieved remarkable results in the promotion and application of independent supercomputing. He made outstanding contributions to my country's supercomputing system's first-ever world-leading position and its six consecutive times achieving the world's number one ranking. He led the Tianhe supercomputing innovation application team to win the first "National Outstanding Engineer Team" title, the first National Innovation and Advancement Award, one National Science and Technology Progress Special Prize, one National Science and Technology Progress First Prize, and seven Military Science and Technology Progress First Prizes.
Conclusions
First, it is worth noting how this data could be used by someone working in the same field. Of course, the GalaxyEDS simulation files are unlikely to run outside their native environment (and if a copy of this application package were found somewhere, one can be sure there would be surprises from the MSS), but at a minimum, from the full 10-petabyte leak one could certainly extract ready-made models, blueprints of various interesting things, detailed material properties, approaches to the most accurate modeling of various processes, etc.
But the primary value here, of course, is that these documents reveal the Chinese approach to the task of developing combat AI models.
Certainly, an important secondary objective of building this simulation pipeline is to provide the Chinese defense industry with a convenient tool for numerically modeling the interaction of their munitions with the adversary's equipment and vice versa, but the stated main goal is specifically the creation of thousands of data points for training an AI that could instantly provide an accurate assessment of the nature and probability of damage to any targets with any munitions under any conditions.
Many were impressed by fragmentary reports of Claude being used in Trump's recent strike on Iran, and those who heard about it can imagine what a massive multiplier of tactical decision making effectiveness an AI would gain by having a trained physics model, capable of accurately and instantly estimating the probabilities and degrees of damage to friendly or enemy targets in each individual case.
If this pilot project succeeds, it will inevitably lead to putting such a synthetic data pipeline on a permanent footing for the continuous generation of synthetic data as the need arises to account for new combinations of equipment, munitions, and conditions of their employment.
It is also worth stating the obvious thought that the choice of targets here unambiguously points to a future war for Taiwan - an American aircraft carrier, HIMARS (whose arsenals Taiwan is constantly expanding), and various bunkers.
Coming up next...
From the leaked 20 GB sample, another large chunk of data worth examining is the stealth theme - specifically, a software package for the simultaneous optimization of RCS and aerodynamic characteristics of aircraft.
On this topic, the leak contains no explanatory documents, but there is a pile of configuration and executable files, from which it may well be possible to at least create some visualizations, and at best - reproduce the operation of this application package.
This will clearly require more time and effort, so I will write about it later.