Blogs / AI in Science: Discovering New Theories and Laws Without Human Intervention

AI in Science: Discovering New Theories and Laws Without Human Intervention

هوش مصنوعی در علم: کشف نظریه‌ها و قوانین جدید بدون دخالت انسان

Introduction

Science has always been built on human curiosity and the ability to recognize patterns. From the moment Newton saw the apple fall to the instance Einstein discovered relativity, humans have been at the center of scientific discovery. But today we are witnessing an unprecedented transformation: artificial intelligence is becoming an independent scientist that can discover new theories, find laws of nature, and create novel materials without human intervention.
This transformation is not a distant dream, but a reality unfolding in leading laboratories worldwide. Autonomous AI systems can now execute the entire research cycle from literature review to experiment design, data analysis, and even writing scientific papers. They have predicted the structure of millions of proteins, discovered new materials that would have taken humans decades to find, and are fundamentally changing how scientific research is conducted.
In this article, we delve deep into this phenomenon: how AI has become an automated scientist, what remarkable achievements it has made so far, what challenges lie ahead for this technology, and what the future of science will look like in the era of automated discovery.

Evolution from AI4S to Agentic Science: Transition from Tool to Scientist

The history of using artificial intelligence in science can be divided into three key stages, each representing a major leap in the capabilities of this technology.

Stage One: AI as a Computational Tool

Initially, AI was merely a tool to accelerate computations. Scientists used machine learning algorithms to analyze big data, identify patterns, and predict outcomes. At this stage, humans still made all important decisions and AI was just a computational assistant.

Stage Two: AI for Science (AI4S)

The second stage began with the emergence of more complex deep learning models. At this stage, AI could perform more specialized tasks such as predicting protein structures, simulating molecules, and discovering complex patterns in scientific data. AI4S demonstrated that artificial intelligence could perform better than humans in certain specific scientific tasks.

Stage Three: Agentic Science - The Era of Artificial Scientists

We have now entered the third stage, called Agentic Science. In this phase, AI systems become autonomous agents that can conduct the entire research process without human supervision. These systems can formulate their own hypotheses, design and execute experiments, analyze results, and even write scientific papers.
Recent research shows that this evolution from assistant to independent research partner is one of the most important scientific advances of the last decade, which can dramatically increase the speed of scientific discovery.

Architecture of Autonomous Scientific Discovery Systems

Autonomous scientific discovery systems typically consist of several key modules that work together:

1. Literature Review and Knowledge Extraction Module

This module uses large language models like GPT or Claude to read, summarize, and extract key knowledge from thousands of scientific papers. Unlike humans who take months to review a field's literature, these systems can do this in hours.

2. Hypothesis Generation Module

Using extracted knowledge, the system generates new hypotheses. These hypotheses may be combinations that no human has thought of. Neural network algorithms and reinforcement learning help the system identify promising hypotheses.

3. Experiment Design and Execution Module

After generating a hypothesis, the system must design an experiment to test it. In automated laboratories, robots independently execute experiments. This module can conduct hundreds of experiments simultaneously using AI and robotics.

4. Data Analysis and Inference Module

Data from experiments are examined using data analysis algorithms and unsupervised learning. The system identifies patterns and concludes whether the hypothesis has been confirmed or not.

5. Writing and Publishing Module

Some advanced systems can even write scientific papers and prepare them for publication. Of course, this part still requires human review.
This architecture is implemented using multi-agent systems and frameworks like LangChain, CrewAI, and AutoGen.

AlphaFold: The Turning Point in Autonomous Scientific Discovery

One of the greatest successes of AI in scientific discovery is AlphaFold, developed by Google DeepMind. This system managed to solve one of biology's 50-year challenges: predicting the three-dimensional structure of proteins knowing only their amino acid sequences.

Why is Protein Structure Prediction Important?

Proteins are the fundamental building blocks of life, and understanding their three-dimensional structure is essential for comprehending their function. Before AlphaFold, determining a protein's structure could take months or even years and cost enormous sums.

AlphaFold's Achievements

AlphaFold, using deep neural networks and advanced supervised learning techniques, has been able to:
  • Predict the structure of over 200 million proteins with atomic accuracy
  • Include nearly all known proteins in its database
  • Reduce a process that took years to minutes
  • Achieve accuracy beyond all previous methods

AlphaFold 3: The Next Generation

The third version of AlphaFold has gone even further and can predict interactions of proteins with DNA, RNA, ligands, and ions. This advancement could revolutionize new drug discovery and understanding diseases.
AlphaFold demonstrated that AI can not only help scientists but can solve problems that decades of human effort couldn't address.

Discovering New Materials: From Years to Days

Another field where AI has revolutionized is discovering and designing novel materials. Traditionally, discovering a new material could take 10 to 20 years and cost millions of dollars.

GNoME: Discovering Millions of New Materials

Google DeepMind has developed a system called Graph Networks for Materials Exploration (GNoME) that can autonomously discover new materials. This system:
  • Has predicted 2.2 million new crystalline structures
  • Identified 380,000 stable materials that are likely to be synthesizable in laboratories
  • 736 of these materials have been created in laboratories by independent researchers and confirmed to exist

Autonomous Laboratories: 10 Times Faster

Recently, autonomous laboratories have been built that can completely independently discover new materials. These laboratories, using reinforcement learning and advanced robotics:
  • Collect data 10 times faster than traditional methods
  • Conduct dynamic and real-time experiments instead of static ones
  • Execute the entire discovery cycle without human intervention
Recent research shows that combining physics laws with AI (Physics-informed AI) can discover new materials even with limited data. This approach allows systems to use existing physical knowledge to guide their search.

Fully Automated Research Systems: AI-Researcher and Similar Examples

One of the most advanced autonomous scientific discovery systems is AI-Researcher, which was published in a paper in May. This system demonstrates that the entire research cycle can be completely automated.

AI-Researcher's Capabilities

This system includes the following modules:
  • Automatic literature review: Reading and analyzing thousands of scientific papers
  • Hypothesis generation: Creating new hypotheses based on gaps in existing knowledge
  • Algorithm design: Writing code and designing computational methods to test hypotheses
  • Experiment execution: Conducting simulations and computational experiments
  • Results analysis: Examining data and evaluating statistical significance
  • Paper writing: Writing complete scientific reports with charts and references

The Nobel Turing Challenge: Can AI Win a Nobel Prize?

A question recently raised is: Can AI make a discovery worthy of winning a Nobel Prize? Some researchers predict this may happen soon.
According to the Nobel Turing Challenge definition, for a scientific discovery to be attributed to AI, it must be:
  • Completely or largely autonomous
  • Made without direct human intervention in the discovery process
  • Have significant scientific importance
Some experts believe AlphaFold itself could have been a candidate for such an award, although human supervision and intervention still existed in its development process.

Discovering Laws of Nature: From Data to Equations

One of the most fascinating applications of AI in science is automatic discovery of physical laws. Systems have been built that can observe data and extract the mathematical equations governing natural phenomena.

Symbolic Regression: Discovering Equations from Data

Symbolic Regression algorithms can extract mathematical equations from experimental data. Unlike traditional methods that only optimize parameters of a specified equation, these algorithms discover the equation itself.

Successful Examples

  • Rediscovering classical physics laws like Newton's law of gravitation and laws of motion solely from experimental data
  • Finding new relationships in scientific data that humans hadn't noticed
  • Accelerating the mathematical modeling process of complex phenomena
This capability shows that AI can not only analyze data but can discover the fundamental laws governing nature.

Current Challenges and Limitations

Despite remarkable progress, serious challenges still exist on the path to fully automated scientific discovery:

1. Limitation in True Creativity

Current AI primarily works based on existing patterns. Its ability to create completely new conceptual leaps is still limited. Revolutionary discoveries like general relativity that require fundamental rethinking of concepts are still beyond the capabilities of current systems.

2. Explainability and Interpretation

One serious problem is result interpretability. When AI makes a discovery, it often cannot explain why this discovery is correct. This issue relates to explainable AI and is important for scientific acceptance.

3. Computational Costs

Training and running large language models and complex AI systems requires enormous computational resources. These costs can be barriers to widespread access to these technologies.

4. Validation and Verification

Every scientific discovery needs independent validation. Reviewing and confirming discoveries made by AI still requires careful human oversight to ensure their correctness. The issue of AI hallucination is a serious concern.

5. Ethical Considerations

Important questions about ethics in AI arise:
  • Who owns AI discoveries?
  • Should AI's name be listed as an author on papers?
  • How can we prevent misuse of this technology?

Key Technologies Behind Autonomous Scientific Discovery

The success of automated scientific discovery systems is due to the combination of several advanced technologies:

1. Transformer Models

Transformer models are the foundation of modern large language models. The attention mechanism in these models allows them to understand complex relationships in data.

2. Specialized Neural Networks

3. Novel Architectures

4. Deep Learning Frameworks

Powerful tools like TensorFlow, PyTorch, and Keras have made the development of these systems possible.

5. Quantum Computing and AI

Combining quantum computing and AI could take scientific discovery capabilities to a new level. Quantum computing can solve complex problems impossible for classical computers.

The Future of Scientific Discovery: From Collaboration to Complete Autonomy

The future relationship between humans and AI in science can be envisioned in several ways:

Scenario 1: Augmented Intelligence

In this scenario, AI works as an extremely capable partner alongside scientists. AI handles repetitive tasks, massive data analysis, and complex simulations, while humans provide creativity, philosophical insight, and ethical judgment.

Scenario 2: Automated Discovery with Human Oversight

In this model, AI systems discover autonomously, but humans play the role of observer, guide, and validator. This is likely the most probable scenario for the near future.

Scenario 3: Autonomous AI Scientific Agents

In this advanced scenario, AI systems become completely autonomous and can execute research projects from start to finish without any human intervention. This scenario raises serious ethical and philosophical challenges.

Scenario 4: AGI and Science

If Artificial General Intelligence (AGI) is achieved, we might witness an explosion of scientific progress. An AGI could work in all scientific fields simultaneously and discover complex interdisciplinary relationships that no human could see. The topic of life after AGI emergence is one of the most important futuristic discussions.

Practical Applications of Automated Scientific Discovery in Various Fields

1. Medicine and Drug Discovery

  • New drug discovery: AI can examine millions of molecular combinations in a fraction of the time
  • Personalized medicine design: Analyzing individuals' genomes to design specific treatments
  • Predicting side effects: Identifying potential problems before clinical trials

2. Materials Science and Engineering

  • Discovering advanced batteries for energy storage
  • Designing superconducting materials at higher temperatures
  • Developing more efficient catalysts for chemical reactions

3. Physics and Astronomy

  • Analyzing telescope data to discover new planets
  • Simulating the first moments of the universe
  • Discovering new fundamental particles in particle collisions

4. Smart Agriculture

  • Optimizing agricultural products using genetics
  • Predicting plant diseases
  • Designing optimal irrigation systems

5. Environmental Science and Climate Change

  • More accurate predictive modeling of climate change
  • Discovering new carbon capture methods
  • Optimizing renewable energy

The Role of Advanced Language Models in Scientific Discovery

New language models play a very important role in scientific discovery:

GPT-5 and the Next Generation

GPT-5 and advanced reasoning models like O3 Mini and O4 Mini have better capabilities for complex scientific reasoning. These models use Chain of Thought techniques to produce more logical reasoning.

Claude Opus 4.1 and Claude Sonnet 4.5

Claude Opus 4.1 and Claude Sonnet 4 also perform exceptionally well in analyzing scientific texts and assisting research. Claude Sonnet 4.5, as Anthropic's smartest model, has advanced capabilities in processing complex information.

Gemini and DeepSeek

Google's Gemini 2.5 Flash and DeepSeek are also powerful tools for scientific analysis. DeepSeek V3.2 has high efficiency using Sparse Attention.

Grok 4 and AI Model Competition

xAI's Grok 4 also shows significant performance as one of the advanced chat models in AI programming models comparison and GPT-5 vs Claude 4.1 comparison.

Complementary Technologies: Multimodality and Advanced Learning

Multimodal Models

Multimodal models can process text, images, audio, and video simultaneously. This capability is extremely valuable for analyzing complex scientific data.

Learning with Limited Data

Techniques like zero-shot and few-shot learning allow systems to learn new tasks even with limited data.

Optimization and Efficiency

Techniques like LoRA for efficient fine-tuning and AI optimization help reduce computational costs.

Democratization of Scientific Discovery

One of the most important consequences of AI in science is the democratization of access to research tools:

Open-Source Tools

Open-source platforms like Hugging Face allow researchers worldwide to access advanced models.

Google Colab and Cloud Tools

Using Google Colab for training deep learning models makes research possible for those who don't have expensive computational resources. Google Cloud AI tools also help in this area.

Small Language Models

Small Language Models (SLM) that are more efficient can run on personal devices and broaden access to AI.

AI's Role in Predicting the Future of Science

Financial and Economic Predictive Modeling

Predictive financial modeling with AI shows how historical data can be used for future prediction.

Time Series Forecasting

Tools like Prophet and ARIMA are used to predict scientific and technological trends.

Simulation and World Models

World Models can simulate the future and help scientists predict the probable outcomes of their experiments.

Social and Economic Challenges

Impact on Scientific Careers

The impact of AI on jobs is a serious concern. Will human scientists be replaced? Or will their role change?

Inequality in Access

There is a risk that wealthy countries and institutions will have access to more advanced tools and the scientific gap will widen.

Security and Misuse

The impact of AI on cybersecurity shows that this technology can also be used for malicious purposes. Automated scientific discoveries could be misused to create dangerous weapons or harmful materials.

The Future of AI in Science: Beyond Human?

The big question is: Can AI surpass humans in scientific creativity?

Current Limitations

Current language models have limitations. Language models' limitations in understanding human language show that these systems have not yet achieved deep human understanding.

Can AI Become Conscious?

The debate on AI and consciousness is one of the deepest philosophical questions. If AI becomes conscious, are its discoveries truly "scientific"?

Future Developments

Conclusion: Human-Machine Coexistence in Science

Artificial intelligence is fundamentally changing how scientific research is conducted. From discovering the structure of millions of proteins to finding new materials and even generating original hypotheses, this technology has demonstrated amazing capabilities.
But the future of science is not necessarily a future where humans are eliminated. Rather, we will likely witness a dynamic coexistence between human creativity and machine computational power. Humans ask big questions, make ethical judgments, and create meaning, while AI opens new ways to find answers with unparalleled speed and precision.
The main challenge is not resisting this change, but shaping it to ensure that automated scientific discoveries benefit all humanity, not just a select few. With careful planning, ethical adherence, and international cooperation, we can use this golden age of scientific discovery to solve humanity's greatest challenges.
The future of science is a future where humans and machines, hand in hand, expand the boundaries of knowledge into unknown territories.