A presentation by Richard West r.west@northeastern.edu @richardhwest on Dec 6, 2022
to the faculty and staff meeting of the Department of Chemical Engineering at Northeastern University. Discussing (very) recent advances in artificial intelligence and the opportunities and threats it offers to chemical engineering research, scholarship, creativity, education, and assessment.
Double Revolving field theory-how the rotor develops torque
(Very) Recent AI advances for Chemical Engineering research and education
1. (Very) Recent Advances
in Artificial Intelligence
• Opportunities and threats for
Chemical Engineering research and education
Much of this is
breaking news
A presentation by Richard West r.west@northeastern.edu @richardhwest on Dec 6, 2022
to the faculty and staff meeting of the Department of Chemical Engineering at Northeastern
24. Designing degree programs…
• Write two degree program descriptions:
1. The Bachelor of Science in Chemical Engineering and Computer Science
provides expertise in computational modeling and simulation of chemical
processes. The curriculum is designed to prepare students to practice in the
engineering and control of processes involving chemicals, biotechnology
feedstocks, and pharmaceuticals, as well as the fundamentals of program design,
software development, and algorithms and data.
2. The Bachelor of Science in Chemical Engineering and Data Science provides
students with the technical and analytical skills to work with large datasets. The
curriculum focuses on the application of data science methods to solve problems
in the chemical engineering field, including the development of predictive models,
optimization of processes, and the design of experiments. Students will also learn
data visualization, machine learning, and artificial intelligence techniques.
(Generated with GPT3 using a “completion” task. Prompt in white, response in grey)
25. Writing
• I asked GPT3 "Suggest 5 titles for this conference
abstract: <pasted abstract>"
• 1. Automating the Search for Environmentally-Friendly Halocarbon
Alternatives
• 2. Screening for Greener Refrigerants and Suppressants with RMG-
Cantera Modeling
• 3. Exploring the Flammability of Halocarbon Blends with Reaction
Mechanism Generator
• 4. Using Automated Kinetic Models to Predict the Flame Speeds of
Halocarbons
• 5. Discovering New Pathways of Halocarbon Suppression with Reaction
Mechanism Generator
• We used one of them
30. Not easy to get it right
I think with careful rubric design and prompt engineering there may be a way
– but probably only worth the effort for a very large class
31. AI image generation
A database of chemical species and reaction
properties and associated algorithms for
estimating thermodynamic and chemical
kinetic properties relevant for predictive
chemical kinetic models. This database can
be accessed through the Reaction
Mechanism Generator website and directly
through a Python interface. Image made
(by me) with Midjourney and DALL-E AI
image generators. J. Chem. Inf. Model.
2022, 62, 20, 4906–4915
https://doi.org/10.1021/acs.jcim.2c00965
32. Diffusion models have other uses.
Images are made using a
“diffusion model” trained to
gradually remove noise
Things move fast in this field!
(video)
(video)
34. Back to the image generation..
“a chemical
engineering professor
drinking a cup of tea in
a chemistry lab”
35. What do you notice?
Apparently, all chemical engineering professors (who drink tea) are white men.
36. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
37. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
38. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
39. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
40. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
41. Language models also have biases.
They try to hide, mask, or avoid it, but it’s relatively easy to trick it into revealing itself
42. Galactica
• Large language model
specifically for science
• Great at writing grants!
• Launched by Meta
(Facebook) on Nov. 15
• After 3 days, closed to
public because it was
generating very realistic
but fake and dangerous
“science”.
• https://www.technologyreview.com/2022/11/18/
1063487/meta-large-language-model-ai-only-
survived-three-days-gpt-3-science/
44. Other AI tools to try
• Elicit.org
• Uses language models to help literature review.
• Summarizes papers for you. https://elicit.org
• CoPilot (in VSCode)
• Writes software code for you, like predictive text messages
(but for whole functions!) https://github.com/features/copilot
• Lex.page
• AI-powered word processor, to beat writers’ block
• I just got off the waitlist but haven’t tried it yet. https://lex.page/
• (Ask me if you want an invitation)
• OpenAI.org
• https://chat.openai.com/ ChatGPT.
• https://beta.openai.com/ GPT language models.
• Image generation
• Midjourney https://www.midjourney.com
• DALL-E 2 https://openai.com/dall-e-2
45. That’s all, folks…
• “a student sitting at a desk,
using an AI-powered tool
on their laptop to complete
a homework assignment. In
the background, the
chalkboard reads "AI in
Education" and a red pen
hovers nearby, ready to
mark any mistakes. The
image conveys the
potential of AI in
education, as well as the
potential for cheating.”
• (the image prompt was
generated by ChatGPT,
based on the email to
Becky at the start of this
presentation, the images
were made by Midjourney)