--> 23% increase by 2030 in AI related careers in STEM (McKinley 2023)

AI is a way of processing information that depends on large datasets (Wen et al 2022). AI models are able to take in massive amounts of data to make predictions or create results, something chemical experimentation would highly benefit from

• GANS are generative adversarial networks, a type of machine learning that both creates and tests synthetic data creation to make the most accurate results (Feuerriegel et al 2023)

• Usage of AI in chemistry can aid in making numerous data collection methods more efficient, but is widely looked over. 

  • For example, in spectroscopy, AI models can help identify compounds used and the functional groups it belongs to.

• Applications of AI in chemistry would make a lot of basic experimentation more efficient. The ability to run simulations over and over again would also create an abundance of results which would make our overall result a lot more accurate. 

Hybrid intelligence is the combination of human intelligence and AI to complete tasks in a more efficient way. 

Hybrid learning is where AI is used collaboratively with human intelligence. With the creation of an AI model, it is thus important to incorporate elements of human intelligence within it (Feuerriegel et. all 2023).

→AI is modeled after the human brain, many models use “neural networks” mimicking the connections made between neurons of a brain.

→ “there is extensive feedback modulation in human visual information processing. The current deep networks are primarily based upon feedforward signals, although there are efforts to incorporate feedback processing into the network.

AI’s failure to use contextual information/prior knowledge in a feedback loop is a setback that CNNs work to improve, something human brains are able to do easily.

The effectiveness of AI relies on its ability to constantly learn from itself.

A combination of AI models may work in a more human-like manner than an AI model alone. By utilizing AI models and conversing them between each other, it becomes clear that they are able to learn from each other and build upon each other in a very similar way to humans and their collaborative nature.

AI is thus able to make these paths easier and faster to see the results of than if you were to experiment with each path length.

--> Amon Machine Learning experiment, where a target molecule ChemML is a program with machine  learning algorithms tailored to chemistry

• With AI, calculations being done by computers regarding chemical reactions/characteristics, are done a lot faster and with a lot more data backing it. AI is able to constantly learn from training data in a way that no human can possibly accomplish, which is why it makes it so imperative that it is leveraged for completing chemistry experiments 

AI is being increasingly used in the chemistry field, prompting for more human-AI integration in chemical informatics.

The accessibility of open-source AI and machine learning fundamentals such as PyTorch (a python extension) has aided in the growing field of AI in chemistry. 

Many chemists are coming from data-literate backgrounds, making it easier for them to learn the processes of AI. AI in chemistry is proven to automate laboratory practices and make them less effort than before.

The need for chemists to be data-literate is rising, and it is important for them to know how to implement artificial intelligence into their laboratory practices to make them more efficient and more automated.

Chemists can use AI’s ability to optimize reaction conditions, and finding synthetic routes to complex target molecules.

Due to the nature of computing power and technological innovation moving at an increasingly faster pace, it is important for chemists to learn how to utilize AI in their work.

Highlights that LLMs can be used in research writing as it is able to reduce the time taken on tasks such as formatting, making the research more easily replicable and allows the researcher to take more time on producing the best results possible in comparison to focusing on the semantics in the writing.

In mole-to-mole conversion, AI is able to instantly complete these conversions, a lot faster than a human can and a lot more accurately. There are less factors that could lead to human error, which is where a lot of experimental results can go wrong.

AI uses big datasets consistent with chemicals and analyzes them to which will work best for that specific chemical reaction.

It is very important to preprocess data for chemical data to be useful in the usage of artificial intelligence algorithms. Artifacts, or distortions to the data, can set back results and make them less reliable.

AI data generation process isn’t a simple, straightforward process. Instead, there are many steps to its procedures that include a lot of failure within the algorithm.

This is described as the feedback loop. It also stresses the importance of datasets that are updating constantly, and one who correctly aligns to the purpose of the AI model. 

The study revealed that chatGPT sometimes had struggles in identifying chemical rules and the nature of some compounds. It had problems with problem-solving because of arithmetic errors. It was unable to apply chemical concepts to an everyday context. ChatGPT was also unable to translate images (graphical representations)

ChatGPT is not able to process complex chemical structures due to the representation they are in. It was unable to correctly identify the data in the first place, making the predictive accuracy and solutions less accurate. 

Ex. a CNN (discriminative model) only spots a few stark differences between cats and dogs, whereas DCGAN can generate new photos of cats and dog by analyzing ALL components of their pictures (called "features"), is able to use what the CNN may have skipped over

→ Generative models as able to learn correlations within datasets of images in comparison to discriminative models that are simply used to label data that is given to the model. It also gives an example of how this difference can be seen.Generative models are able to use patterns from the overall distribution of data to create more data.

Discriminative models have largely been used in problem solving as many use classifier-based solutions, without considering generative models and their usefulness in this field. Generative models are able to select features and help with classification and model accuracy.

It is first turned into a 2D matrix, and then is used to display 3D information like atomic coordinates/bond angles. Linear notations such as SMILES and InChI are used to transform the chemical’s 3D structure into its symbolic representation, used in the AI to find patterns.

Predictive accuracy is also increased with this preliminary step, as this data representation is the only way the machine learning model can truly take in.

--> Shows how the discriminator and generator parts of the GAN model work to classify probabilities present within a dataset. It explains how his data is represented within the algorithm (1s and 0s)

Identifies the principle (binary cross entropy) that is used as the loss function in the generator. A loss function stops training in the generator if it is not able to “catch up” to the discriminator. It is a complicated equation that can be implemented into the code without much mathematical difficulty as it is just built into the code, but it is important to know what type of loss function is being used and why.