Lab Releases Vacuum Deposition Emulator and 46-Task Benchmark for AI Agents
Testing an AI agent on vacuum deposition equipment has been hard: you need an expensive chamber and an industrial PLC. The lab has released its control software, an emulator that mimics Siemens PLCs, and an MCP server under Apache-2.0. The MCP server is the bridge that lets Claude Code or a local model connect to the software. Everything runs on a Windows laptop, and the ADS-Bench test provides 46 tasks: 40 short ones and 6 end-to-end scenarios. Models can be compared without risking hardware.
AI-processed from Habr AI; edited by Hamidun News
Testing an AI agent that controls a vacuum deposition system for thin films requires the system itself and an industrial PLC, the hardware block that opens valves and switches on pumps. Most teams have neither, and a model error in a real chamber is costly. The lab, which previously reported on X-ray mirror deposition using an LLM and a local Qwen 3.
8 acting as an operator, has released the APS Control Community package under the Apache-2.0 license. It includes a working deposition control program, a hardware emulator that imitates Siemens controllers over the network, and an MCP server.
MCP is the protocol through which a language model or Claude Code gets access to the program's tools. For testing agents, there is an open benchmark, ADS-Bench v2.0: 46 tasks, including 40 short exercises and 6 end-to-end scenarios.
Everything runs on an ordinary Windows PC without administrator rights. The interface is built on WinForms, so the .NET 10 Desktop Runtime is required. The current release is v1.7.2. The installer is not digitally signed yet, so SmartScreen will show a warning on first launch.
The Emulator as a Separate Process
Normally, a debug mode is built into the code itself: conditional compilation switches the program between a real port and a stub. The authors took a different approach. The control program contains no directives such as #if EMULATE. The drivers for Siemens S7 controllers (port 102) and peripheral instruments (ports 5025–5033) always communicate through ordinary TCP sockets. In the lab, the sockets connect to real hardware; on the laptop, they connect to the emulator at the local addresses 127.0.0.1–127.0.0.5.
The practical implication is simple: code you debugged against the emulator runs on the real installation unchanged. The program does not know whether a rack with a PLC or a C# model is answering it.
There is also an engineering detail useful to anyone writing software for hardware. Each communication bus gets its own dedicated persistent thread, without Task.Run or await. If one controller hangs on a network timeout, the others are unaffected, and the operator window does not freeze. A watchdog raises an alarm when the connection drops, so stale readings do not remain on screen.
What the Emulator Calculates and How Recipes Are Written
The emulator calculates the overall gas and heat balance in the chamber. Pumping speed depends quadratically on the throttle valve opening: if you close it to 25%, pumping slows by a factor of 16. Heating of the substrate holder is calculated using the Stefan–Boltzmann law. You do not have to wait for pumping in real time: the model clock runs ten or more times faster. During a run, you can trigger an overheat or a node failure and observe how the protections respond.
The package includes three profiles: a simple chamber with two magnetrons (Sputter), a system with a load lock and four magnetrons (ADS), and a four-chamber cluster with a transfer robot (Cluster). In the cluster profile, the emulator runs five independent virtual PLCs, and scenarios must coordinate pumping and gate valve opening across different controllers.
The process is defined by a text recipe in an XML file with the .rcpx extension. The built-in validate_recipe analyzer catches errors before the run starts. One example from the article: if you close the argon supply immediately without resetting the pressure setpoint, the chamber switches to the Idle standby mode, and the throttle keeps holding 3 mTorr. After 60 seconds, a vacuum alarm fires. The command pwait 0 1 before closing the gas prevents this.
Agent Security and the Benchmark
Network access for agents is disabled by default. The Enable-AgentAccess.ps1 script creates a local server on 127.0.0.1:8026 with TLS and key verification, then prints a ready-made claude mcp add command for connecting the APS.Mcp.exe bridge.
The seventeen MCP tools are split into three levels:
- Read: reading the system schematic and sensor values;
- Recipe: validating and running recipes and exporting logs, enabled by default;
- Manual: direct control of nodes, granted only with the
-Manualflag.
Even with a key, control commands will not go through until a person switches the control panel to Online-Remote mode. In local mode, the model can read sensors and validate recipes, but it cannot open a valve or start deposition. The safety interlocks run at the control-system level and cannot be bypassed from the editor or over the network. For example, the vent valve checks eight conditions, and a 750 Torr pressure interlock prevents high voltage from being applied to the magnetrons in a vented chamber.
Alongside the program, the authors released the aps-operator-skill operator skill (CC-BY-4.0) and the ADS-Bench v2.0 benchmark. It contains 46 tasks across all three profiles, including 22 tasks carried over from the authors' previous article. The local Qwen 3.8 was tested on a graphics card with 16 GB of memory.
Limitations
The emulator does not model plasma or three-dimensional gas dynamics. It does not reproduce gradual target erosion, arcing, or drift in deposition rate over long runs. It is sufficient for debugging interlocks, valve logic, and agent behavior, but it does not replace testing on a real installation. The build runs only on Windows.
For agent developers, the project is interesting beyond deposition. It is a ready-made example of how to give a model access to dangerous equipment: the emulator as a separate process, access levels, mandatory pre-validation of recipes, and physical confirmation by a human. The repository accepts issues and pull requests, including drivers for other hardware.
- Nov 21, 2026Opus rewrote llama.cpp code: Qwen runs 40% faster on a CPU alone
- Nov 21, 2026Jensen Huang Steps Down as NVIDIA CEO After 33 Years, CFO Colette Kress Becomes Interim CEO
- Nov 21, 2026OpenAI Disbands Robotics Team Again: GPT-6 Astra Completed 74.8% of Tasks in Simulation Versus 51%
- Nov 21, 2026AWS gives agents a shared LLM cache skill: up to 40% cheaper, first token 80% faster
Key Questions Answered
What are the key details on The Emulator as a Separate Process?
Normally, a debug mode is built into the code itself: conditional compilation switches the program between a real port and a stub. The authors took a different approach. The control program contains no directives such as #if EMULATE. The drivers for Siemens S7 controllers (port 102) and peripheral instruments (ports 5025–5033) always communicate through ordinary TCP sockets. In the lab, the sockets connect to real hardware; on the laptop, they connect to the emulator at the local addresses 127.0.0.1–127.0.0.5.
What the Emulator Calculates and How Recipes Are Written?
The emulator calculates the overall gas and heat balance in the chamber. Pumping speed depends quadratically on the throttle valve opening: if you close it to 25%, pumping slows by a factor of 16. Heating of the substrate holder is calculated using the Stefan–Boltzmann law. You do not have to wait for pumping in real time: the model clock runs ten or more times faster. During a run, you can trigger an overheat or a node failure and observe how the protections respond.
What are the key details on Agent Security and the Benchmark?
Network access for agents is disabled by default. The Enable-AgentAccess.ps1 script creates a local server on 127.0.0.1:8026 with TLS and key verification, then prints a ready-made claude mcp add command for connecting the APS.Mcp.exe bridge. The seventeen MCP tools are split into three levels: Read: reading the system schematic and sensor values; Recipe: validating and running recipes and exporting logs, enabled by default; Manual: direct control of nodes, granted only with the -Manual flag.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.