Meeting minutes and associated discussions.

[utkarshp1161]

Meeting Minutes - Utkarsh and Austin - 4th March 2026

1. Look at database mode to handle the devices

• Problem - sending dictionaries like below is pain if there are many devices

uv run python -m tango.test_context asyncroscopy.Microscope.Microscope \
  --host 127.0.0.1 \
  --port 8889 \
  --prop "{
    'haadf_device_address': 'tango://127.0.0.1:8888/test/nodb/haadf#dbase=no',
    'autoscript_host_ip': 'localhost',
    'autoscript_host_port': 9090
  }"

---

2. EELS Acquisition and CEOS corrector

• Discussion on whether EELS acquisition should be implemented directly as device with its own acquistition.
• If implemented
  - this way, eels.py will need to include Gatan-specific logic for handling EELS detectors and acquisition.
  - so If EELS is from another manufacturer we need to change logic in the detectors which is undesireable
• Same logic for CEOS

---

3. Logging

Purpose:

• Enable consistent logging for device servers for cases when Microscope is suppose to get say an array but returns an Error string
• Improve debugging and operational monitoring.

Discussed implementing logging using PyTango's recommended approach.
Reference:

• https://tango-controls.readthedocs.io/projects/pytango/en/latest/tutorial/logging.html

---

4. Testing Strategy

• Testing is important to flag:
  • Python version changes
  • PyTango /numpy version changes which leads to data type changes
  • Hardware specific version change leads to modified command- example AS version 1.15 -

Current status :

• test for haadf detector works but Microscope one doesnt
• Probably as it needs the devices running in an specific way?
• Hint : look at - https://tango-controls.readthedocs.io/projects/pytango/en/latest/how-to/testing/testing_approaches.html
• In long term we need it as GitHub actions so tests run for PR's

---

5. Documentation Improvements

• Current documentation is useful for understanding how to:
  • Add detectors
  • Extend the microscope interface

However:

• Documentation needs updates to reflect the new microscope base class (Microscope base abstraction).
• Future contributors should be able to easily extend the framework for new detectors or microscope types.

---

6. Use of ABC for Microscope Abstraction

Question raised:

Is using Python ABC (Abstract Base Classes) useful for abstracting the microscope interface?

Issue identified:

• Combining @detector decorators with @abstractmethod leads to design conflicts.
• The goal is to provide a Tango-compatible blueprint for microscope device servers that others can implement.

Conclusion:

• Using ABC in the current design may not provide practical benefits.
• A simpler blueprint/interface approach may be preferable for extensibility.

[DomPTech]

For number 6., is the @detector decorator something that's being custom making or a part of PyTango? I can't seem to find it in either Tango's source code or asyncroscopy's.

[utkarshp1161]

Ah my bad, its @command in PyTango.

[AustinHouston]

Meeting minutes the next day

Simple digital twin for image generation - @DomPTech

• copy ThermoMicroscope, gut it, nd replace with random np functions for now

Tests @DomPTech

Working on databse mode @utkarshp1161

Conversation with all about restructuring base microscope nd how to include CEOS/Gatan
End result is write them as their own hardware for now, inherit device, and the thermo microscope calls those devices in the helper commands. The reasoning is, we will write it for our microscope and get something going. Anyways, many other TF microscopes integrate with Gatan and CEOS.

Add these functions to incoorperate ceos to microscope proxy @AustinHouston
Make documentation ipynbs for integration with jupyter books @AustinHouston @utkarshp1161
Make Jupyter books @gduscher good for talks
Diagram the new structure flow chart @AustinHouston
How to use ABC for microscope abstraction, nonimplemented error is very good for debugging

[utkarshp1161]

• Handled the DB mode and Microscope abstraction here - feat: DB mode and Microscope abstraction (ABC conflict) #40
• Please comment directly on the PR if any inputs

[utkarshp1161]

Meeting minutes for 13th Mar 2026 - Friday

1. Timeout-error on Windows :

   • Currently we are getting timeout error when we run the Server on the Windows
   • @gduscher will test it on his laptop
   • ✅ Done

2. All notebook should now reflect database mode - @utkarshp1161

3. CEOS server integration to microscope and also Base class.

   • Microscope.get_aberrations -> gives the aberrations
   • @AustinHouston
   • EELS detector will follow same design after this - Microscope.get_spectrum("EELS")

4. More on twins - @AustinHouston and @utkarshp1161

   • Needed for testing drift correction algorithms - Drift correct #25
   • Need to illustrate the case where GD mentioned sending diffraction spots only is sufficient - @AustinHouston and @gduscher
     - probably use abTEM to get diffraction simulation?
     - then send only the spots?
   • Needed for testing aberration correction methods using ML
   • Something which loads a HAADF and SI and we can try methods like
     - Segmentation
     - Inpainting
   • Also a good place to test LLM ? - any thoughts @AustinHouston, @DomPTech and @whittlegears
   • HAADF_EDS twin - 5220108 - 5e7c018 - ✅ Done

5. Design: Needs more thought!:

   • AH pointed that everytime we use detector using mic.get_spectrum/get_image - it checks whether the detector proxy is intialized or not - see here in base class
     • Overkill as everytime you get a signal you are checking whteher detector is there or not --> it makes sense to check it only once at start --> Better to have it in intitialization?
     • Interestingly this came up when we were adding EDS detector - Great catch!
     • check is removed - bea49c3 - ✅ Done -

6. More on tests

   • Need to think! what tests should cover and does current tests have it? - @utkarshp1161 and @DomPTech
     - Ideally it should flag version/datatype/typing changes
     - Python, Numpy, PyTango, AS
     - also borrow inspiration from Pytango tests
   • Need to extend PR tests to windows platform as well @DomPTech
   • Add hooks to copilot or claude - so they can give initial comments - @DomPTech
   • Ease of debugging - Need to cover the case when "if AS server is off on the Microscope computer" a reasonable error is thrown
     • maybe tests is not the good place for this --> ideally the Microscope classes should handle this
   • we have a test for ThermoMicrscope - 70f064b - ✅ Done -

7. Docs

   • GD pointed out to have a good figure for the README - @AustinHouston
   • Doc needed for base microscope
   • This needs to be modified based on current version
   • Needed Jupyter books in the long run - good for workshop talks - @AustinHouston @gduscher

Extras:

• Good time to make PyTango - main branch - ✅ Done

• Need a good name (eg: smart proxy?) as GD pointed out

• Release v-0.1.0

• Maybe later - GUI over the tango servers which is used by LLM - @whittlegears and @DomPTech

[DomPTech]

I also had a thought about the Microscope base class. Right now we have STEM-specific commands (beam blanking, acquiring STEM images, etc.) in this class, but if we want to add AFM control this would not make sense. I'm thinking we should create an abstract subclass of Microscope called STEMMicroscope and then ThermoMicroscope inherits STEMMicroscope.

[AustinHouston]

on point 5, we will move the check to the init and make it more useful for debugging

[AustinHouston]

also, we (@gduscher and I) solved the windows error! (point 1).
Main point was setting environement variable TANGO_HOST = :9094 under system properties advanced tab

[utkarshp1161]

Timeout-error on Windows :

is resolved --> GD, AH and UP tested both cases

• Microscope-pc as server and Microscope-pc as client
• Microscope-pc as server and Mac-laptop as client

Verified - Only one port(Tango db port) access is sufficient
Also a nice thing about this is client doesn't need "manufacturer" python modules (eg: AS wheels)

[AustinHouston]

to set the system properties search for sysdm.cpl

[AustinHouston]

It's time to make pytango main branch, definitely

[utkarshp1161]

Currently I can think of this way to do it:

• current main branch is copied on another branch for later reference say twisted-based.
• PyTango replaces the current main branch

[utkarshp1161]

Handled this:
PyTango branch is now main. The previous Twisted-based implementation is preserved in the twisted-legacy branch for reference.

[utkarshp1161]

Lets maintain PyTango branch for a week until PR's(#47) get resolved.

[gduscher]

Hi Domique, I am afraid that will lead to too many subclasses. In my mind the base class should be microscope and that inherits special microscope specific comands in ThermoMicroscope ... We need more discussion, Gerd

…

________________________________ From: DomPTech ***@***.***> Sent: Monday, March 16, 2026 8:29 AM To: pycroscopy/asyncroscopy ***@***.***> Cc: Duscher, Gerd ***@***.***>; Mention ***@***.***> Subject: Re: [pycroscopy/asyncroscopy] Meeting minutes and associated discussions. (Discussion #37) I also had a thought about the Microscope base class. Right now we have STEM-specific commands (beam blanking, acquiring STEM images, etc.) in this class, but if we want to add AFM control this would not make sense. I'm thinking we should create an abstract subclass of Microscope called STEMMicroscope and then ThermoMicroscope inherits STEMMicroscope. — Reply to this email directly, view it on GitHub<#37 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AL53B4CGXNABCPQ2BT5OO2D4Q7XRFAVCNFSM6AAAAACWHGEFCWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTMMJVHA3TONQ>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[AustinHouston]

base should not have init defined
think about the decide properties at the top of base
remove adorned logic on base -> move to thermo
twins - need drift twin

tests - don't push if the tests are failing
reach out

[AustinHouston]

think about making it private for now

naming
we are based on PyTango
is there a better name for the project?

[utkarshp1161]

Meeting Minutes

1. PR and Testing

• Discussed PRs as a useful tool to keep everyone aligned and enable effective code review.
• Emphasized that tests are important to catch undesirable situations early and avoid time-consuming black-box debugging.
  • Example: When a collaborator runs the project (e.g., on Windows 8 or another environment), tests help diagnose issues.
• Local developers should run tests before pushing changes.
  • If tests fail, consult the person who wrote the code.

Writing More Tests

• Refer to the following for inspiration:
  • https://gitlab.com/tango-controls/pytango/-/tree/develop/tests
  • https://github.com/bluesky/tiled/tree/main/tests
  • https://github.com/bluesky/ophyd-async/tree/main/tests/system_tests_tango

---

2. Design

• Remove device attributes from the base Microscope.
• Remove adorned image logic from the base Microscope.
• Client notebooks can be placed in a separate public repository.

---

3. Open Questions

• How should someone add a new detector or a twin microscope?
  • It becomes challenging to decide on what patterns to pick - given we have base Microscope and inherited Microscope class like - ThermoMicroscope
  • Cases where the twin commands should directly run also with ThermoMicroscope is important
  • Write docs here - https://github.com/pycroscopy/asyncroscopy/tree/main/docs

[DomPTech]

MCP Server Discussion:

• The MCP server will run on top of Tango in DB mode.
• It will autofind all functions and wrap them in tools.
• We can specify a blocklist of tools we don't want the Agent to have access to.

[utkarshp1161]

Discussion today(including Px-hackathons) - 10th April 2026 - Friday:

1. ML PC - Utkarsh shares requirements and everyone edits

   • use case:
     - Storage : accessed though tiled servers
     - GPU for prototyping workflows

2. MCP server :
   - try goose as mcp client
   - supports chat ui to interact
   - plug mcp servers with a button
   - write pyTEMlib mcp server:
   - example : https://github.com/pycroscopy/SciFiReaders/tree/main/SciFiReaders/mcp
   - need to make pytemlib functions more specific : eg: read_file as function name is very generic and might confuse the LLM

3. Moving away from DASK as it is slow(GD pointed out):

   • Preference:
     - data provenance - The complete history of where the data came from and how it was created, processed, and modified over time.
     - GPU native

4. Marimo over jupyter notebooks:

• they have all features of notebook but are light weight .py files
• can lead to data provenance solution?
• the can be served on a url easily and can serve as a interactive tutorial

[utkarshp1161]

Proposed Machine Learning Workstation (~$5,000 Budget)

Component	Recommendation	Specs / Details	Estimated Cost (USD)	Notes
GPU	NVIDIA RTX 4090	24GB VRAM	$1,800 – $2,000	Primary ML compute (PyTorch, CUDA, DL workloads)
CPU	AMD Ryzen 9 7950X	16 cores / 32 threads	$500 – $600	Strong for data preprocessing + multitasking
Motherboard	X670E / B650E (ATX)	PCIe 5.0, multiple M.2 slots	$250 – $350	Future expandability (GPU + storage)
RAM	128GB DDR5	4 × 32GB	$400 – $600	Large datasets, embeddings, microscopy data
Primary Storage	2TB NVMe SSD (Gen4)	~7000 MB/s	$150 – $250	OS + active datasets
Secondary Storage	4TB SSD (NVMe/SATA)	—	$250 – $400	Dataset storage
Power Supply (PSU)	1000W–1200W Gold	Fully modular	$180 – $250	Required for RTX 4090 stability
Cooling	360mm AIO Liquid Cooler	—	$120 – $200	Sustained CPU workloads
Case	ATX Mid/Full Tower	High airflow	$100 – $200	Thermal performance + space
Networking	Onboard (2.5GbE/WiFi 6E)	—	Included	Fast data transfer
Operating System	Ubuntu Linux	LTS version	$0	Standard ML ecosystem

| Total Estimated Cost | | | $3,800 – $4,800 | Leaves buffer for peripherals / price fluctuations |

[whittlegears]

I have a few notes:

The AMD Ryzen 9 9950X is $545 and has better performance and is around the same price.
1000W–1200W Gold with the native 12VHPWR connector is a good idea so the adapter won't melt.
Let's get the X670E specifically (not B650E) for the motherboard.
We also might want to consider getting a different motherboard (like an x870e) because we might want to expand to additional nicer GPUs in the future like a 5090.
Also it looks like the 4090 might be difficult to get as it is discontinued but I'm assuming we don't want to switch to the 5080 and don't have the budget for the 5090. Otherwise everything else looks great.

[utkarshp1161]

Update on the PC -> I entered the details in an excel sheet and shared with Gerd today.

[utkarshp1161]

Meeting Minutes — 17 April 2026

---

1. Asyncroscopy Package figure plan

---

2. Best LLM Benchmark Discussion

The question posed: "If you had the best LLM, what would you ask it to do?" — framed as a benchmark for testing on Twin.

Current asyncroscopy capabilities:

1. List available microscopes
2. List available devices
3. List available detectors
4. Set microscope to STEM mode and acquire an HAADF image
5. Align the gun lens by adjusting screen value to maximum
6. Acquire atomic-resolution HAADF image — adjusting FOV, focus, and astigmatism as needed
7. Acquire EDS spectra on a 9×9 grid centred on the HAADF image at 300 pA beam current

Near-term goals:

1. Acquire an image — check microscope mode first, then query which detector to use
2. Acquire EELS on a 9×9 grid at 10 pA beam current
3. List available image detectors
4. Set convergence angle to 10 mrad
5. Acquire a convergent beam electron diffraction (CBED) pattern with the Ceta camera
6. Ensure the main screen is down whenever possible
7. Report current microscope state

Future functionality:

1. Tilt to zone axis
2. General tilt control

---

3. Other MCP Server Functionalities

• pyTEMlib
• SciFiReaders — MCP server already exists repo
• Segmentation — could be exposed as a device
• Atom finding AI — fast, simple; assumes known material, rotation, and FOV

---

4. Projects

• Random atom distribution: spacing 20–200 pm, intensity 0.1–1.0

5. Also what context would be good to include in data?

---

6. MBR Meeting (2–3 pm)

• Goal to integrate agentic systems with AFM/STEM/synthesis
  • Set up a weekly Teams call — Wednesdays 9 am - done
    • Attendees: AFM, STEM, and Synthesis teams
    • Agenda: share current API state and plan next development steps

---

Resources - to try before we meet next time

• MCP client worth trying: Goose — https://goose-docs.ai/
• Good MCP reference: https://gofastmcp.com/getting-started/welcome
• Try the SciFiReaders MCP server ([repo](https://github.com/pycroscopy/SciFiReaders/tree/main/SciFiReaders/mcp)) inside Goose

[AustinHouston]

Action Items

Gerd

MCP for pyTEMlib
send benchmark questions

Levi

Documentation

Dom

Notebook for benchmarking prompts

###Utkarsh
Add EELS detector

Austin

Add Ceta Camera
Working with Darel on UNet blob finder

Darel

Test UNet blob finder idea

[utkarshp1161]

Meeting minutes - Thursday, 28 May 2025 — @utkarshp1161 and @AustinHouston

Q. What is the PR of interest ?
PR: #83

Q. Are there TODO's added?

• Review all TODOs in the repo: grep -rn "# TODO" .

Q. How to handle List_of_adorned_images?

• We get it From AutoScript → next step is how to interface reading from Tiled
• TBD
• Done in - Dev ach #91

Q. Philosophy of asyncroscopy wrt users(Academia or Industry(Context: Utkarsh at covalent))

• Smart acquisition, not high-throughput
• Tiled saves time; TIFF workflows remain compatible

Q. Why Documentation should be kept up?

• Current docs - https://github.com/pycroscopy/asyncroscopy/tree/main/docs
• We as developers want to add features 2 weeks later we forget so it is a mental cue
• Guide for users if they suggest a feature and want to modify behaviour
• Contributor onboarding (where to add features)
• What to add:
  • Documentation of - https://github.com/pycroscopy/asyncroscopy/blob/main/asyncroscopy/software/DATA.py

Testing

• Add test for - DataServer (Tiled) integration — uses AutoScript from Gatan PC which is a twin but the datastrucutres like adornedimages can be obtained from it and entire pipeline can be run. Very handy when version changes to dependencies(PyTango, Tiled and more)

LLM use discussion - To be followed up:

• Idea : Reading code is very important to not loose the control to LLM's
  - https://www.heise.de/en/blog/Reading-code-instead-of-writing-code-The-underestimated-senior-discipline-11295136.html
• Definition of reading needs to be re-thought as we can just not use LLM's blindly.
• Why I like to use LLM(MIndful use: as costs on both pocket and environment)
  - We can ask it to look at unfamiliar api ( Like Hardware api) and then suggest how it can be used
  - It can suggest a plan and then we can agree or disagree and modify the plan --> after the plan we can choose to code manually
  - Good at giving code reviews
• One idea is to have prompts as part of PR review process. So one reads the prompt and then reads the code. It kind of helps the reviewer the context of what the PR tries to solve.
• The PR comment can have also the prompts which the developer used to write the code