Source | Xinzhiyuan ID | AI-era
After AI is combined with various scientific fields, a technological revolution full of potential and challenges is taking place.
By exploring theories, designing experiments, and analyzing data, AI will supercharge scientific discovery as we know it.
On August 2, researchers from the Google team published a study in Nature—Scientific Discovery in the Era of Artificial Intelligence, which summarized the application and progress of AI in scientific discovery.
Paper address: https://www.nature.com/articles/s41586-023-06221-2
How data is collected, transformed, and understood provides the basis for developing scientific insights and theories.
The emergence of deep learning in the early 2010s has greatly expanded the scope and ambition of these scientific discovery processes.
Artificial intelligence is increasingly being applied across scientific disciplines to integrate massive data sets, refine measurements, guide experiments, explore theoretical spaces that match data, and provide actionable and reliable models that integrate with scientific workflows, thereby Achieve self-discovery.
Data collection and analysis are fundamental to scientific understanding and discovery, and are two central goals of science, quantitative methods and emerging technologies.
The introduction of digitization in the 1950s paved the way for the widespread use of computers in scientific research.
Since the 2010s, the rise of data science has enabled AI to identify scientifically relevant patterns from large datasets, thereby providing valuable guidance.
Although scientific practices and processes vary across stages of scientific research, the development of AI algorithms spans traditionally siled disciplines.
Such algorithms can enhance the design and execution of scientific studies and are becoming an indispensable tool for researchers.
Scientific discovery is a multifaceted process involving several interrelated stages, including hypothesis formation, experimental design, data acquisition, and analysis
Recently, the latest advances in AI in science, including unraveling the protein folding problem 50 years ago, and the simulation of molecular systems with millions of particles driven by artificial intelligence, have proved the potential of artificial intelligence to solve challenging scientific problems.
As with any new technology, the success of AI4Science depends on our ability to integrate it into our daily practice and understand its potential and limitations.
Barriers to widespread adoption of AI in the scientific discovery process include internal and external factors specific to each stage of the discovery process, and concerns about the usefulness and potential misuse of methods, theories, software, and hardware.
In the paper, the researchers will explore the development of artificial intelligence science and address key questions.
AI-assisted scientific research data collection and management
The increasing size and complexity of data sets collected by experimental platforms has led to scientific research increasingly relying on real-time processing and high-performance computing to selectively store and analyze data generated at high speed.
data selection
A typical particle collision experiment generates over 100 terabytes of data per second. Such scientific experiments are pushing the limits of existing data transmission and storage technologies.
In these physics experiments, more than 99.99% of the raw instrument data are background events that must be detected and discarded in real time to manage the data rate.
To identify rare events for future scientific research, the deep learning approach replaces pre-programmed hardware event triggers with algorithms that "search for outliers" to detect unexpected or rare phenomena that may have been missed during the compression process.
The background process can use a deep autoencoder to generate a model.
Autoencoders return high loss values (anomaly scores) for previously unseen signals that do not belong to the background distribution (rare events). Unlike supervised anomaly detection, unsupervised anomaly detection does not require annotations and has been widely used in physics, neuroscience, earth science, oceanography, and astronomy.
Data annotation
Training a supervised model requires a dataset with annotations that provide supervised information to guide model training and estimate the function or conditional distribution of the target variable from the input.
In biology, techniques for assigning functional and structural labels to newly characterized molecules are critical for downstream training of supervised models, as experimentally generating labels is very difficult.
For example, despite advances in next-generation sequencing technologies, less than 1% of sequenced proteins have been annotated with biological function.
Another data labeling strategy is to use an agent model trained on human-labeled data to label unlabeled samples, and use these predicted pseudo-labels to supervise downstream predictive models.
In contrast, label propagation diffuses labels into unlabeled samples via a similarity graph constructed based on feature embeddings.
In addition to automatic labeling, active learning can also identify the most informative data points to be manually labeled or the most informative experiments to be performed.
In this way, the model can be trained with fewer labels provided by experts. Another strategy for data labeling is to use domain knowledge to formulate labeling rules.
data generation
The performance of deep learning continues to improve as the quality, diversity, and size of training datasets improve.
An effective way to create better models is to augment training datasets by generating additional synthetic data points through automatic data augmentation and deep generative models.
In addition to manually designing such data augmentation, reinforcement learning methods can discover an automatic data augmentation strategy that is both flexible and independent of downstream models.
Deep generative models, including variational autoencoders, generative adversarial networks, normalized flow, and diffusion models, can learn the underlying data distribution and sample training points from an optimized distribution.
Generative adversarial networks have proven useful for scientific imagery as they can synthesize realistic images in many domains.
Probabilistic programming is an emerging technique in generative models, and expresses generative models of data as computer programs.
Learning meaningful representations for scientific data
Deep learning can extract meaningful representations of scientific data at different levels of abstraction and optimize them to guide research, usually through end-to-end learning.
A high-quality representation should retain as much information as possible from the data while keeping it simple and understandable.
Scientifically meaningful representations should be compact, discriminative, able to discriminate between underlying variables, and encode underlying mechanisms that can be generalized across multiple tasks.
Here, the researchers introduce 3 emerging strategies to meet these requirements: geometric priors, self-supervised learning, and language modeling.
geometric prior
Since geometry and structure play a central role in scientific domains, integrating "geometric priors" in learning representations has been shown to be effective.
Symmetry is a widely studied concept in geometry. It can describe the behavior of mathematical functions in terms of invariance and arithmetic differences to represent the behavior of neural feature encoders under a set of transformations.
In scientific image analysis, objects do not change as they are translated in the image, which means that image segmentation masks are translationally equivariant because they change equivalently when input pixels are translated.
Incorporating symmetry into a model can benefit AI in limited annotated data by increasing the training samples and can improve extrapolated predictions for inputs that are significantly different from those encountered during model training.
Geometry Deep Learning
Graph neural networks, have emerged as a dominant approach for deep learning on datasets with underlying geometric and relational structures.
Broadly speaking, geometric deep learning involves discovering relational patterns and encoding local information in the form of graphs and sets of transformations through neural information transfer algorithms.
Learning meaningful representations of scientific data
self-supervised learning
Supervised learning may not be sufficient when only a few labeled samples are available for model training, or when the cost of labeling data for a particular task is prohibitive.
In this case, leveraging both labeled and unlabeled data can improve model performance and learning.
Self-supervised learning is a technique that enables a model to learn general characteristics of a dataset without relying on explicit labels.
Self-supervised learning is an important preprocessing step to learn transferable features on large unlabeled datasets, and then fine-tune the model on small labeled datasets to perform downstream tasks.
This pre-trained model with broad knowledge of the scientific domain is a general-purpose predictor that can be applied to a variety of tasks, thereby improving annotation efficiency and surpassing purely supervised methods.
language modeling
Masked language modeling is a popular method for self-supervised learning of natural language and biological sequences.
As natural language and biological sequence processing continue to evolve, they inform each other's development.
During training, the goal is to predict the next token in the sequence, while in mask-based training, the self-supervised task is to recover the masked token in the sequence using bidirectional sequence context.
Protein language models can encode amino acid sequences to capture structural and functional properties and assess the evolutionary fitness of viral variants.
Transformer architecture
Transformers are neural architecture models that can process token sequences by flexibly simulating interactions between arbitrary token pairs, surpassing earlier efforts to model sequences using recurrent neural networks.
Although Transformers unify graph neural networks and language models, the running time and memory footprint of Transformers may scale quadratically with the sequence length, resulting in long-range modeling, and linearized attention mechanisms are challenged in terms of efficiency.
Therefore, unsupervised or self-supervised generative pre-training transformers are widely used, followed by parameter-efficient fine-tuning.
neural operator
Standard neural network models may not be adequate for scientific applications because they assume that the data dispersion is fixed.
This approach is not applicable to many scientific datasets collected at different resolutions, and grids.
Furthermore, data are often sampled from underlying physical phenomena in the continuous domain,
Neural operators learn representations that are not affected by discretization by learning mappings between function spaces.
Neural operators are discretization invariant, which means they can handle any discretized input and converge to a limit during mesh refinement.
Once a neural operator is trained, it can be evaluated at any resolution without retraining. In contrast, the performance of standard neural networks degrades when the data resolution during deployment changes from the data resolution at model training time.
AI-based scientific hypothesis generation
Testable hypotheses are at the heart of scientific discovery.
Black-box predictors of scientific hypotheses
Identifying promising hypotheses for scientific inquiry requires efficiently examining many candidate scenarios and selecting those that maximize the output of downstream simulations and experiments.
In drug discovery, high-throughput screening can evaluate thousands to millions of molecules, and algorithms can prioritize which molecules to study experimentally. Models can be trained to predict the utility of experiments, such as relevant molecular properties, or symbolic formulas that fit observations.
However, for many molecules, experimentally factual data for these predictors may not be available.
Therefore, weakly supervised learning methods can be used to train these models, where noisy, limited or imprecise supervision is used as the training signal.
These methods can cost-effectively replace annotations by human experts, expensive computations in silicon, or higher-fidelity experiments.
AI-guided scientific hypothesis generation
AI methods trained on high-fidelity simulations have been used to efficiently screen large molecular libraries.
To further improve the efficiency of these processes, AI-selected candidates can be fed to medium or low-throughput experiments for continuous refinement of the candidates using experiment feedback.
Results can be fed back into AI models using active learning and Bayesian optimization, enabling the algorithm to improve its predictions and focus on the most promising candidates.
AI methods become very valuable when hypotheses involve complex objects such as molecules.
For example, in terms of protein folding, AlphaFold2 can predict the three-dimensional atomic coordinates of proteins based on amino acid sequences, and its accuracy can even reach the atomic level, even if the protein structure is different from any protein in the training data set.
This breakthrough has led to the development of various AI-driven protein folding methods, such as RoseTTAFold106.
In addition to forward problems, artificial intelligence methods are also increasingly used in inverse problems, aiming to understand the causal factors that produce a set of observations.
Inverse problems, such as inverse folding or fixed backbone design, can use black-box predictors trained on millions of protein structures to predict amino acid sequences from the three-dimensional atomic coordinates of protein backbones.
However, such black-box AI predictors require large training datasets and, while reducing reliance on existing scientific knowledge, have limited interpretability.
Navigating Combinatorial Hypothesis Spaces
Although sampling all the hypotheses that fit the data is daunting, finding a good one is a manageable goal, which can be formulated as an optimization problem.
Compared with traditional methods that rely on human-designed rules, artificial intelligence strategies can be used to estimate the return of each search and prioritize search directions with higher value.
Agents typically trained with reinforcement learning algorithms to learn policies.
The agent learns to take actions in the search space that maximize a reward signal, which can be defined to reflect the quality of the generated hypotheses or other relevant criteria.
To solve optimization problems, evolutionary algorithms can be used to solve symbolic regression tasks. The algorithm generates random sign laws as an initial solution set.
In each generation, the candidate solutions change slightly.
The algorithm checks whether any modification produces a signed law that fits the observations better than previous solutions, and the best solution is retained for the next generation.
However, reinforcement learning methods are gradually replacing this standard strategy.
Reinforcement learning uses neural networks to sequentially generate mathematical expressions by adding mathematical symbols from a predefined vocabulary and using a learned policy to decide which symbol to add next.
A mathematical formula is represented as a parse tree. The learning strategy takes as input a parse tree to decide which leaf node to expand and which symbol to add.
Another way to use neural networks to solve math problems is to convert mathematical formulas into sequences of binary symbols.
The neural network strategy can then increment the binary characters one at a time in order of probability.
By devising a reward to measure the ability to refute a conjecture, this method can find ways to refute a mathematical conjecture without prior knowledge of the mathematical problem.
Combinatorial optimization is also applicable to tasks such as discovering molecules with desirable drug properties, where each step in molecular design is a discrete decision-making process.
In this process, a partially generated molecular graph is given as input to a learning policy that makes discrete choices about where to add new atoms and which atoms to add at selected positions in the molecule.
By iteratively performing this process, the strategy generates a range of possible molecular structures, evaluated based on their suitability for target properties.
The AI agent learns policies that anticipate actions that initially seem unorthodox, but turn out to be effective.
In mathematics, for example, supervised models can identify patterns and relationships between mathematical objects and help guide intuition and generate conjectures.
These analyzes point to previously unknown patterns, or even new models of the world.
However, reinforcement learning methods may not generalize well to unseen data during model training because once the agent finds a sequence of actions that works well, it may get stuck in local optima.
To improve generalization, some exploration strategies are needed to collect a wider range of search trajectories that can help the agent perform better in new and modified settings.
Optimizing Differentiable Hypothesis Spaces
Scientific hypotheses often take the form of discrete objects, such as symbolic formulas in physics or chemical compounds in pharmaceutical and materials science.
Although combinatorial optimization techniques have been successful on some of these problems, differentiable spaces can also be used for optimization as it lends itself to gradient-based methods that can efficiently find local optima.
To be able to use gradient-based optimization methods, two methods are often used.
The first is to use models such as VAEs, which map discrete candidate hypotheses to points in the latent variable space.
The second approach is to relax the discrete assumptions into differentiable objects that can be optimized in a differentiable space.
This relaxation can take different forms, such as replacing discrete variables with continuous variables, or using soft versions of the original constraints.
Symbolic regression applications in physics use the syntax VAE. These models represent discrete symbolic expressions as parse trees using context-free grammars and map the parse trees into differentiable latent spaces.
Bayesian optimization is then employed to optimize the sign-law latent space while ensuring that the expressions are syntactically valid.
In many scientific disciplines, the space of hypotheses can be much larger than what experiments can investigate.
Therefore, we urgently need a method to efficiently search and identify high-quality candidate solutions in these largely unexplored regions.
AI-Driven Experiments and Simulations
Assessing scientific hypotheses experimentally is critical to scientific discovery.
However, lab experiments can be prohibitively expensive and impractical.
Computer simulations have emerged as a promising alternative, with the advantage of being more efficient and flexible than experiments.
While simulations rely on hand-crafted parameters and pioneering approaches to simulate real-world scenarios, compared to physical experiments, there are trade-offs between accuracy and speed that require an understanding of the underlying mechanisms.
However, with the advent of deep learning, these challenges are being addressed by identifying and optimizing hypotheses for efficient testing and by giving computer simulations the ability to connect observations to hypotheses.
Efficiently evaluate scientific hypotheses
AI systems provide experimental design and optimization tools that can enhance traditional scientific methods, reducing the number of experiments required and conserving resources.
Specifically, the AI system could assist with two important steps of experimental testing: planning and guidance.
In traditional methods, these steps often require trial and error, which can be inefficient, expensive, and sometimes even life-threatening.
The AI Initiative provides a systematic approach to designing experiments, optimizing their efficiency, and exploring uncharted territory.
At the same time, AI guidance steers the experimental process toward high-yielding hypotheses, enabling the system to learn from previous observations and adjust the experimental process.
These AI methods can be based on simulation and prior knowledge for model building, or based on pure machine learning algorithms for model building.
AI systems can assist experiment planning by optimizing resource usage and reducing unnecessary investigations. Unlike hypothesis searching, experimental planning involves the procedures and steps involved in the design of scientific experiments.
An example is a chemical synthesis program. Chemical synthesis planning involves finding the sequence of steps by which a target compound can be synthesized from an existing compound.
AI systems can design synthetic pathways to obtain desired compounds, reducing the need for human intervention.
Active learning has also been used for materials discovery and synthesis. Active learning involves iterative interaction with experimental feedback to refine hypotheses. Materials synthesis is a complex and resource-intensive process that requires efficient exploration of high-dimensional parameter spaces.
Active learning exploits uncertainty estimation to explore the parameter space and reduce the uncertainty as little as possible.
Decisions often need to be adjusted in real time during the course of an experiment. However, this process can be difficult or error-prone when relying solely on human experience and intuition. Reinforcement learning offers an alternative approach to continuously respond to changing environments and maximize the safety and guaranteed success of experiments.
For example, in the experiment of magnetron tokamak plasma, the reinforcement learning method interacts with the tokamak simulator to optimize the strategy of the control process (as shown in the figure below).
In another study, a reinforcement learning agent steered a stratospheric balloon based on real-time feedback, such as wind speed and sun altitude, and looked for favorable wind currents for navigation.
In quantum physics, experimental design needs to be dynamically adjusted based on the best choice for future complex experiments, and reinforcement learning methods can overcome this problem by iteratively designing experiments and obtaining feedback from them.
For example, reinforcement learning algorithms have been used to optimize the measurement and control of quantum systems, thereby increasing experimental efficiency and accuracy.
Using Simulations to Deduce Observations from Hypotheses
Computer simulations are a powerful tool for deriving observations from hypotheses, enabling the evaluation of hypotheses that are not directly testable.
However, existing simulation techniques rely heavily on human understanding and knowledge of the underlying mechanisms of the system under study, which may make simulations less than optimal and efficient.
AI systems can enhance computer simulations by learning more accurately and efficiently, better fitting key parameters of complex systems, solving differential equations governing complex systems, and modeling the state of complex systems.
Scientists typically study complex systems by creating models involving parametric forms, which require domain-specific knowledge to identify initial symbolic expressions for the parameters.
For example, molecular force fields are interpretable but limited in their ability to represent various functions and require strong inductive biases or scientific knowledge to generate.
To improve the accuracy of molecular simulations, an AI-based neural potential fit to expensive but accurate quantum mechanical data has been developed, replacing conventional force fields.
Furthermore, uncertainty quantification has been used to localize energy barriers in high-dimensional free energy surfaces, thereby improving the efficiency of molecular dynamics169 (below).
For coarse-grained molecular dynamics, AI models can reduce the computational cost of large systems by determining how coarse-grained the system needs to be from the learned hidden complex structure.
In quantum physics, neural networks have replaced manually estimated symbolic forms of wave functions or density functionals due to their flexibility and ability to fit data accurately.
Differential equations are crucial for modeling the dynamics of complex systems in space and time. AI-based neural solvers blend data and physics more seamlessly than numerical algebraic solvers.
These neural solvers combine physics with the flexibility of deep learning by modeling neural networks based on domain knowledge (below).
AI methods have been applied to differential equation solving in various fields, including computational fluid dynamics, predicting the structure of glass systems, solving difficult chemical kinetic problems, and solving Eikonal equations to characterize the travel time of seismic waves.
In dynamical modeling, regular differential equations can be used to model continuous time. Neural networks can be parameterized in the spatio-temporal domain by physical information loss to the solution of the Navier-Stokes equations.
However, standard convolutional neural networks have limited ability to model fine-grained features of solutions. This problem can be solved by learning operators that model mappings between functions with neural networks.
Furthermore, the solver must be able to adapt to different domains and boundary conditions, which can be achieved through graph partitioning by combining neural differential equations with graph neural networks.
Statistical modeling is a powerful tool that can provide a comprehensive quantitative description of complex systems by modeling the distribution of states in them.
Normalized Flow can use a series of reversible neural networks to map any complex distribution to a prior distribution (such as a simple Gaussian distribution), and vice versa.
While computationally expensive (typically requiring hundreds or thousands of neural layers), normalized flow provides an accurate density function, enabling sampling and training.
Unlike conventional simulations, normalized flows can generate equilibrium states by directly sampling from prior distributions and applying neural networks, such that the computational cost is fixed.
This enhances sampling in lattice field and gauge theory, improving Markov chain Monte Carlo methods that might otherwise fail to converge due to modal mixing.
major challenge
To take advantage of scientific data, models must build on human expertise and then use simulations to enhance model performance.
This integration opens up new opportunities for scientific discovery.
However, to further enhance the impact of AI in science, major advances in theory, methodology, software, and hardware infrastructure are required.
Collaboration across disciplines is critical to achieving a comprehensive and practical approach to advancing science through AI.
practical considerations
Scientific datasets are often poor candidates for AI due to limitations in measurement techniques that produce incomplete datasets, biased or conflicting readings, and limited data accessibility due to privacy and security concerns analyze.
Standardized and transparent formats are needed to ease the workload of data processing.
Model cards and data sheets are examples of efforts to document the operational properties of scientific datasets and models.
Additionally, federated learning and encryption algorithms can be used to prevent the public release of sensitive data with high commercial value into the public domain.
Leveraging open scientific literature, natural language processing and knowledge graph technologies can facilitate literature mining, contributing to the advancement of materials discovery, chemical synthesis, and therapeutic science.
The use of deep learning presents complex challenges for human-involved AI-driven design, discovery, and evaluation.
To automate scientific workflows, optimize large-scale simulation codes, and operate instruments, autonomous robotic control can exploit predictions and run experiments on high-throughput synthesis and test lines, creating autonomous laboratories.
Early applications of generative models in materials exploration have shown that millions of possible materials with desired properties and functions can be identified and their synthesizability assessed.
For example, King et al. combined logical AI and robotics to autonomously generate functional genomics hypotheses about yeast and used laboratory automation to experimentally test these hypotheses.
In chemical synthesis, AI optimizes candidate synthetic pathways, and then robots guide chemical reactions based on the predicted synthetic pathways.
Implementing an AI system involves complex software and hardware engineering, requiring a series of interdependent steps, from data screening and processing to algorithm implementation and user application interface design.
Small differences in implementation can lead to significant changes in performance and affect the success of integrating AI models into scientific practice.
Therefore, standardization of data and models needs to be considered. Due to the random nature of model training, variations in model parameters, and changing training datasets, AI methods can have reproducibility issues that are both data- and task-dependent.
Standardized benchmarks and experimental designs can alleviate these problems. Another direction for improving reproducibility is through open source initiatives that publish open models, datasets, and educational projects.
algorithm innovation
In order to contribute to scientific understanding or to acquire scientific understanding autonomously, algorithmic innovation is required to establish an underlying ecosystem of using optimal algorithms throughout the scientific process.
The problem of generalization beyond distributions is at the forefront of AI research.
A neural network trained on a certain range of data may discover patterns that do not apply to a different range of data because the underlying distribution of the latter has shifted.
While many scientific laws are not universally applicable, they generally have broad applicability. And the human brain can adapt to modified environments better and faster than state-of-the-art AI.
There is a very interesting hypothesis that says that humans not only build statistical models based on what they observe, but also build a causal model.
This is a collection of statistical models indexed by all possible interventions (e.g. different initial states, different agent behaviors, or different situations).
Incorporating causality into AI is still an unstudied field, and much work remains to be done.
Techniques such as self-supervised learning have great potential for scientific problems because they can take advantage of large amounts of unlabeled data and transfer the knowledge contained in it to low-data domains.
However, current transfer learning schemes may be temporary solutions in specific situations, lack theoretical guidance, and are vulnerable to changes in the underlying distribution.
While some initial attempts have addressed this challenge, further exploration is needed to systematically measure transferability across domains and prevent negative transfer.
Furthermore, to address the difficulties scientists care about, the development and evaluation of AI methods must be carried out in real-world situations, such as possible synthetic pathways in drug design, and include well-calibrated inaccuracies before transferring them to practical applications. Deterministic estimation to assess the reliability of the model.
Scientific data are multimodal and include images (e.g. black hole images in cosmology), natural language (e.g. scientific literature), time series (e.g. thermal yellowing of materials), sequences (e.g. biological sequences), graphs (e.g. complex systems) and structures (e.g. 3D protein-ligand conformations).
AI methods often operate as black boxes, meaning that users do not fully understand how the output is generated, and which inputs are critical in generating the output.
Black-box models can reduce user trust in predictions and have limited application in domains where model outputs must be understood before they can be realized, such as in human space exploration, and in domains where predictions guide policy, such as in climate science.
Despite a plethora of interpretation techniques, transparent deep learning models remain elusive.
However, the human brain is capable of synthesizing high-level explanations that can convince other humans, if not perfectly.
This offers hope that by simulating phenomena at a similarly high level of abstraction, future AI models will provide explanations and understandings that are at least as valuable as those provided by the human brain.
This also suggests that studying high-level cognition may inspire future deep learning models that combine current deep learning capabilities with the ability to manipulate articulate abstractions, causal inference, and generalization beyond distributions.
The Impact of AI on the Enterprise of Scientific Research
Looking ahead, the demand for AI expertise will be influenced by two forces.
First, there are areas that could immediately benefit from AI applications, such as autonomous laboratories.
Second, smart tools can advance the state of the art and create new opportunities, such as studies related to the length and timescales of biological, chemical or physical processes that cannot be observed experimentally.
Based on these two forces, we expect changes in the composition of research teams to include AI experts, software and hardware engineers, and new forms of collaboration involving all levels of government, educational institutions, and companies.
Recent state-of-the-art deep learning models keep growing 10,234. These models consist of millions or even billions of parameters and grow tenfold in size every year.
Training these models involves passing data through complex parameterized mathematical operations, with parameters updated to push the model output towards the desired value.
However, the computational and data requirements to compute these updates are enormous, resulting in huge energy consumption and high computational costs.
As a result, large technology companies have invested heavily in computing infrastructure and cloud services, pushing the limits of scale and efficiency.
While for-profit and non-academic organizations have large-scale computing infrastructures, institutions of higher education may be better positioned to integrate across disciplines.
Additionally, academic institutions often have unique historical databases and measurement techniques that may not exist elsewhere but are necessary for AI4Science.
These complementary assets facilitate novel models of industry-university collaboration, which can influence the chosen research questions.
As AI systems approach and surpass human performance, it becomes feasible to use them as a replacement for routine laboratory work.
This approach enables researchers to develop predictive models from experimental data and select experiments to improve those models without manually performing tedious and repetitive tasks.
To support this paradigm shift, educational programs are emerging to train scientists in the design, implementation, and application of laboratory automation and AI in scientific research. These programs help scientists understand when using AI is appropriate and prevent misinterpretation of AI analysis.
in conclusion
AI systems can contribute to scientific understanding, enabling us to study processes and objects that would otherwise not be visualized or probed, and to systematically spark creativity by building models from data combined with simulation and scalable computation.
In order to realize this potential, the safety concerns posed by the use of AI must be addressed through responsible and thoughtful deployment of the technology.
To use AI responsibly in scientific research, scientific research needs to determine the level of uncertainty, error, and utility of AI systems.
This understanding is critical to accurately interpreting AI output and ensuring we do not rely too heavily on potentially flawed results.
As AI systems continue to evolve, prioritizing reliable implementation with appropriate safeguards is key to minimizing risks and maximizing benefits.
AI has the potential to reveal previously unreachable scientific discoveries.
References:
https://www.nature.com/articles/s41586-023-06221-2