A microwave radio frequency chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical guided depth reinforcement learning

Abstract

The application discloses a microwave radio frequency chip multi-port crosstalk self-suppression device and method based on programmable package cavity boundary and physical guided deep reinforcement learning. The device sets programmable boundary units in the package cover plate, side wall, interlayer periphery or core particle area, collects target channel transmission, non-target coupling and package modal response, and constructs a multi-port modal state vector. A physical guided deep reinforcement learning module outputs a boundary coding matrix according to port isolation, target channel maintenance, modal suppression and additional loss constraints, drives the boundary unit to change the equivalent surface impedance and reflection phase, so that the non-target coupling field and the dominant cavity mode form a destructive reconstruction in the disturbed port, while maintaining the performance of the target channel. The scheme is suitable for dynamic electromagnetic regulation and crosstalk suppression of microwave radio frequency chips, Chiplets and 2.5D / 3D advanced packages.

Description

Technical Field

[0001] This invention belongs to the fields of microwave RF chip packaging, advanced packaging electromagnetic compatibility and intelligent electromagnetic control technology. In particular, it relates to a microwave RF chip multi-port crosstalk self-suppression device and method based on a programmable packaging cavity boundary and physical-guided deep reinforcement learning mechanism. It is mainly used for packaging cavity mode control, multi-port crosstalk suppression, target channel maintenance and dynamic electromagnetic boundary reconstruction in microwave RF chips, millimeter-wave chips, chiplet heterogeneous integration, 2.5D / 3D advanced packaging and multi-chip RF systems. Background Technology

[0002] As microwave RF chips, high-density heterogeneous integration, Chiplet, and 2.5D / 3D advanced packaging technologies are all in a phase of rapid development, and the application demands for high frequency, high integration, multi-port parallel operation, and high electromagnetic isolation are increasing, traditional packaging electromagnetic compatibility methods that rely on fixed metal shielding, grounding structures, absorbing materials, fixed electromagnetic bandgap structures, or single matching networks have encountered new and prominent bottlenecks in terms of multi-band adaptability, dynamic control capabilities, and crosstalk suppression under complex task conditions. Specifically, when multiple RF chips, millimeter-wave front-ends, local oscillator chips, power amplifier chips, and control chips are integrated into the same package cavity, complex near-field coupling, spatial leakage, multipath propagation, and cavity intrinsic mode excitation will form between different RF ports, interconnects, chips, and metal boundaries. This can easily lead to problems such as multi-port crosstalk, decreased port isolation, increased target channel insertion loss, group delay fluctuations, and reduced system stability. Even if fixed shielding, periodic structures, or impedance matching methods are used for suppression, the structural parameters usually remain unchanged after the design and manufacturing are completed. Therefore, it is difficult to adapt to dynamic changes in operating frequency, chip activation combination, port operating state, and electromagnetic field distribution inside the package.

[0003] In microwave RF packaging systems, metal covers, package sidewalls, interposers, redistribution layers (RDLs), microbumps, and multi-port RF interconnects naturally and rationally construct packaging spaces suitable for complex electromagnetic propagation. More importantly, non-target coupling within the package is not formed by a single path, but rather by the superposition of direct crosstalk, cavity modes, sidewall reflections, cover reflections, and multipath propagation. Therefore, by altering the local equivalent surface impedance, reflection phase, and electromagnetic propagation conditions of the package boundary, the amplitude and phase relationships of different propagation paths can be reorganized, causing the non-target coupling field to cancel out near the disturbed port, thus achieving active suppression of multi-port crosstalk. Unlike traditional fixed structures, programmable package electromagnetic boundaries can dynamically change their state according to the current operating frequency band, chip combination, and port tasks, thus possessing stronger multi-task adaptability and electromagnetic field reconfiguration capabilities. However, it must be acknowledged that most microwave RF packages currently still use fixed metal boundaries, fixed EBG structures, absorbing materials, or local matching compensation methods. Their control objects, adjustment dimensions, and adaptability are all significantly limited. It remains a major challenge to achieve cavity mode suppression, non-target coupling reduction, and target channel maintenance under multi-band, multi-core, and multi-port conditions in the same packaging platform.

[0004] With the increasing demand for millimeter-wave RF front-ends, chiplet heterogeneous integration, high-speed multi-chip systems, and high-density advanced packaging applications, single fixed boundaries or single matching parameters are no longer sufficient to meet the needs of dynamic tasks, high isolation, and high reliability applications. How to introduce programmable package cavity boundaries into the same microwave RF chip packaging platform, and adaptively generate boundary coding matrices based on different operating frequency bands, chip activation combinations, port states, and dominant package modes, to actively suppress non-target coupling fields and cavity modes while maintaining the transmission performance of the target RF channel, has become a pressing problem in the field of microwave RF packaging and intelligent electromagnetic control. Therefore, it is necessary to propose a microwave RF chip multi-port crosstalk self-suppression device and method based on programmable package cavity boundaries and physically guided deep reinforcement learning, to overcome the limitations of existing fixed-package electromagnetic control technologies in terms of dynamic adaptability, multi-task response capability, modal control dimensions, and intelligent decision-making, and to provide a new technical solution for building a highly isolated, adaptive, reconfigurable, and robust intelligent microwave RF packaging platform. Summary of the Invention

[0005] The purpose of this invention is to propose a self-suppression device and method for multi-port crosstalk of microwave radio frequency chips based on a programmable package cavity boundary and a physical-guided deep reinforcement learning mechanism. This addresses the problem that existing fixed metal boundaries, fixed periodic suppression structures, and single matching compensation methods are difficult to adapt to changes in operating frequency bands, chip activation combinations, and dynamic port tasks, and are also difficult to achieve joint suppression of non-target coupling and dominant package cavity modes while maintaining the performance of the target channel.

[0006] The objective of this invention is achieved as follows: First, multiple programmable boundary units are set at at least two different locations on the inner surface of the metal cover plate, the package sidewall, the periphery of the interposer, the periphery of the redistribution layer, and the non-signal region between the cores of the microwave RF multi-core package. This constructs a programmable package electromagnetic boundary array with multiple switchable equivalent surface impedance states. A multi-port microwave excitation and signal acquisition module is used to acquire the target channel transmission response, non-target port coupling response, and package modal response under different operating frequency bands, core activation combinations, and port operating states, forming a multi-port detection structure that can dynamically sense the electromagnetic propagation state inside the package. Based on the interaction between the microwave electromagnetic field, the intrinsic modes of the package cavity, the coupling paths between cores, and the programmable boundary states, an overall package electromagnetic response model is established. The coupling amounts at different ports, the participation coefficient of the dominant package mode, the target channel transmission performance, the operating frequency, the port activation state, and the current boundary encoding state are constructed into a multi-port modal state vector.

[0007] The multi-port modal state vector is then input into the physics-guided deep reinforcement learning control module. Through the iterative relationship between the current state, boundary coding actions, and control benefits, a boundary coding matrix adapted to the current package operation task is output. This boundary coding matrix controls the impedance state, reflection phase, local loss, or on / off state of programmable boundary cells at different locations, thereby altering the phase relationship and energy distribution between different propagation paths within the package. By coordinating the adjustment of direct crosstalk paths, package cavity modal paths, sidewall reflection paths, and programmable boundary reconstruction paths, the non-target coupling field is destructively coupled near the disturbed port. Simultaneously, the target channel insertion loss, worst-case port isolation, group delay fluctuations, number of boundary switching operations, and additional losses are controlled to meet preset conditions.

[0008] Furthermore, by changing the operating frequency band, chip activation combination, port operating state, dominant mode participation coefficient, boundary impedance state, reward function weight, and physical constraint parameters, different package electromagnetic boundary reconstruction modes can be formed under different task conditions. When non-target port coupling characteristics dominate, the boundary coding state that improves the worst port isolation is prioritized; when the participation of the dominant cavity mode increases, the boundary reconstruction state that reduces the energy of non-target modes is prioritized; when multiple factors such as multi-port coupling, target channel retention, group delay, and additional losses work together, the device forms a multi-parameter jointly optimized electromagnetic boundary control result. Further, by introducing minimum transmission constraints for the target channel, maximum boundary switching quantity constraints, additional insertion loss constraints, device bias range constraints, and package temperature rise constraints, the boundary control process maintains stability under conditions of manufacturing size deviations, dielectric parameter changes, temperature drift, device state errors, and measurement noise. This achieves multi-port crosstalk self-suppression, dominant mode reduction, target channel retention, and dynamic electromagnetic field reconstruction functions based on deep reinforcement learning of programmable package cavity boundaries and physical guidance.

[0009] The present invention may also include: 1. The microwave RF chip multi-port crosstalk self-suppression device of the present invention mainly consists of a microwave RF multi-chip package body, a programmable package electromagnetic boundary array, a multi-port state acquisition module, a modal state reconstruction module, a physically guided deep reinforcement learning control module, and a boundary unit driving module. It is integrated onto the package cover, package sidewall, interposer, redistribution layer (RDL), or chiplet heterogeneous integration platform. Dynamic electromagnetic boundary control can be achieved without changing the basic transmission structure of the original target RF channel. It has a compact structure, strong versatility, and is suitable for microwave RF chips, millimeter-wave chips, chiplets, and 2.5D / 3D advanced packaging applications. Each programmable boundary unit consists of one or more of the following basic execution modules: PIN diode, variable capacitor, RF switch, RFMEMS, ferroelectric device, phase change device, adjustable inductor-capacitor, or other adjustable impedance structures.

[0010] 2. Since different programmable boundary units are distributed on the inner surface of the metal cover, the package sidewall, the periphery of the interposer, the periphery of the redistribution layer, or the inter-core region, each boundary unit can independently set the impedance state, reflection phase, local loss, and control weight. Therefore, dynamic package electromagnetic boundaries under multiple regions, multiple states, and multiple tasks can be naturally and reasonably formed. Different boundary regions are used to realize cavity mode adjustment, inter-core space leakage coupling suppression, direct crosstalk reduction, and multipath propagation path reconstruction.

[0011] 3. Because it uses multi-port state acquisition, modal state reconstruction and physical-guided deep reinforcement learning to achieve dynamic control of the packaged electromagnetic boundary, it can automatically generate the corresponding boundary coding matrix according to the operating frequency band, core activation combination and port task changes. There is no need to redesign the fixed shielding structure or matching network for each operating state. Therefore, it is naturally beneficial to improve the packaged electromagnetic control efficiency, multi-task adaptability and system integration.

[0012] 4. Because the device can form a stable low-coupling, low-modal participation and target channel holding state in the multi-port modal state space under the control conditions, the control results have strong tolerance to packaging processing size deviations, material dielectric parameter fluctuations, device impedance errors, temperature drift and measurement noise. It can still maintain stable crosstalk suppression and electromagnetic boundary reconstruction performance under different operating frequency bands, chip combinations and multi-port parallel operation conditions. Therefore, the requirements for packaging manufacturing consistency and operating environment stability are significantly reduced.

[0013] Since this invention is compared with the prior art, its advantages naturally and appropriately arise: 1. Due to the introduction of programmable package cavity boundaries, multi-port modal state reconstruction, and physical-guided deep reinforcement learning control mechanisms in microwave RF chip packaging, the designed device can dynamically generate corresponding boundary coding matrices for different operating frequency bands, chip activation combinations, and port tasks. This enables the non-target coupling field and the dominant package cavity mode to form controlled cancellation or energy reduction, while maintaining the target RF channel to meet preset transmission requirements. Therefore, compared with fixed metal boundaries, fixed EBG structures, or single matching compensation methods, it greatly improves multi-task adaptability and electromagnetic control dimensions. It also naturally overcomes the various limitations encountered by traditional packaging electromagnetic suppression methods in terms of dynamic reconfigurability and adaptation to complex operating states, and has good flexibility in configuring different packaging structures and operating tasks.

[0014] 2. Because the overall device adopts a modular design that integrates a programmable boundary array, a packaged modal state reconstruction module, and a physically guided deep reinforcement learning control module, it can naturally and reasonably utilize the boundary conditions and multipath propagation characteristics of the package itself to achieve non-target coupling active cancellation and dynamic electromagnetic field reconstruction. Therefore, it does not need to rely on large-area absorbing materials, complex external shielding structures, or fixed matching networks designed for a single frequency point. The structure is compact and conducive to integration with microwave RF chips, millimeter-wave front-ends, chiplet heterogeneous integration, and 2.5D / 3D advanced packaging platforms. This also greatly improves the port isolation capability, dynamic adaptability, and engineering application compatibility in high-density multi-chip RF systems. Attached Figure Description

[0015] Figure 1A schematic diagram of the overall structure of a microwave radio frequency chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical-guided deep reinforcement learning. Figure 2 Multiport RF response and package mode reconfiguration characteristics under different programmable package boundary states; Figure 3 The result of intelligent reconstruction of encapsulated boundaries based on multi-port modal state space and physically guided deep reinforcement learning; Figure 4 Multiport crosstalk suppression and RF response reconfiguration characteristics under different packaged electromagnetic boundary control schemes; Figure 5 Comparison of multi-port modal state distribution and overall control performance under different packaging boundary control schemes; Figure 6 Statistical graph of robustness of multi-port crosstalk self-suppression system of microwave RF chip under random perturbation conditions. Detailed Implementation

[0016] The following is in conjunction with the appendix Figure 1-6 To illustrate the invention in more detail, the following examples are provided: Example 1: This example proposes a multi-port crosstalk self-suppression device and method for microwave RF chips based on programmable package cavity boundaries and physically guided deep reinforcement learning. The device is integrated on a multi-core microwave RF packaging and programmable electromagnetic boundary integrated platform. The packaging interlayer material is a Rogers RO4350B dielectric substrate with a dielectric constant of [missing information]. The substrate thickness is approximately 0.508 mm. The package body comprises two RF functional chips and one local oscillator or control chip, each chip measuring approximately 5 mm × 5 mm. RF signal interconnection is achieved through one or more structures selected from the redistribution layer (RDL), microbumps, bonding wires, or package microstrip transmission lines. The corresponding operating frequency band is 20–50 GHz, with a preferred analysis frequency band of 28–40 GHz. The inner surface of the package metal cover has 8 × 8 programmable boundary impedance units, and auxiliary programmable boundary units are provided in at least one region of the package sidewalls and the peripheral junction area of ​​the interposer. Each boundary unit has at least two equivalent surface impedance states, which are switched using one or more of the following: PIN diodes, RF MEMS, variable capacitors, or RF switches.

[0017] In one specific implementation, the programmable boundary basic unit can be constructed using a metal patch, a slotted metal patch, an open resonant structure, or a local LC loading structure. This embodiment preferably uses a combination of a square metal patch and a central slot, with a unit period of approximately 2.0 mm, a metal patch side length of approximately 1.6 mm, and a gap between adjacent units of approximately 0.2 mm. A PIN diode is connected across both ends of the central slot. The on and off states of the PIN diode correspond to the first and second equivalent surface impedance states, respectively, thereby generating different local reflection amplitudes and phases within the 28–40 GHz target frequency band. When using 2-bit encoding, each basic unit can also form four encoding states (00, 01, 10, and 11) through two independent switching devices to obtain four different equivalent impedances or reflection phases. Each programmable boundary unit is connected to the boundary unit driving module via independent control lines or grouped control lines. The bias control lines are preferably arranged along low-electric-field regions, ground regions, or package sidewalls, and their additional impact on the target RF channel and package cavity modes is reduced through RF choke structures, isolation resistors, or high-impedance bias lines.

[0018] The multi-port RF package is configured with four main detection and operating ports, and simultaneously extracts the target channel transmission response, non-target port coupling response, dominant package mode participation factor, and current boundary coding state. The basic system parameters are: input microwave power approximately 0 dBm, sweep step approximately 20 MHz, 64 programmable boundary elements, each element having 2 or 4 impedance states, internal cover plate height approximately 2.0 mm, chip spacing approximately 3.0–8.0 mm, and detection temperature controlled at 25.0 ± 0.5 ℃. The package operation states are divided into three categories: single-RF channel operation, dual-channel parallel operation, and multi-chip collaborative operation. Different combinations of chips, frequency bands, and ports are used to achieve package electromagnetic field reconstruction and crosstalk suppression under multi-task conditions.

[0019] To specifically implement physically-guided deep reinforcement learning control, this embodiment defines the multi-port modal state vector at time step i as follows: Wherein, represents the current operating frequency. This represents the current port and chip activation state matrix, the coupling state matrix between different non-target ports, and the normalized participation coefficient vector of one or more dominant package cavity modes. Let represent the programmable boundary coding matrix at time i. For a case containing *n* programmable boundary cells, where each boundary cell has *n* selectable impedance states, the reinforcement learning action is... Used to determine the target state among a number of candidate impedance states for at least one programmable boundary unit to switch from its current impedance state, and to generate the boundary encoding matrix for the next time step accordingly. Here, represents the mapping relationship between the current boundary encoding state and the reinforcement learning action to complete the boundary state update.

[0020] The physical-guided deep reinforcement learning control module can employ deep reinforcement learning algorithms based on policy optimization, value function optimization, or Actor-Critic structures. This embodiment preferably uses a policy-optimized deep reinforcement learning network with multi-port modal state vectors. As network input, after state feature mapping is completed through at least two fully connected hidden layers, the output is the probability distribution of the corresponding boundary-encoded action or the boundary control action. The hidden layers can be configured with 128 and 64 neurons respectively, and employ the ReLU activation function. For the 64 programmable boundary units, when each unit has 2 or 4 impedance states, unit-independent encoding, group encoding, or hierarchical encoding can be used to compress the action space, thereby reducing the dimensionality of the action space caused by large-scale boundary encoding combinations and improving control efficiency.

[0021] The reinforcement learning reward function is jointly constructed based on the worst-case port isolation, target channel preservation performance, non-target mode energy, group delay fluctuation, additional insertion loss, and boundary unit switching number, and can be expressed as follows: in, This represents the worst port isolation among all relevant non-target port combinations in the current operating state. Indicates the target channel transmission hold-up index. This represents the normalized energy or modal participation level of one or more dominant non-target encapsulation cavity modes. This indicates the group delay fluctuations within the target frequency band. This represents the additional insertion loss introduced after performing boundary reconstruction. This indicates the number of programmable boundary elements that undergo state changes during a single control process. to This represents the non-negative weighting coefficient of the corresponding performance metric. By adjusting these weighting coefficients, the system can formulate boundary reconfiguration strategies with the primary objectives of improving port isolation, reducing dominant packaging mode energy, maintaining target channel performance, or achieving comprehensive optimization of multiple metrics.

[0022] To further constrain the reinforcement learning control process, the boundary encoding optimization can also satisfy... as well as in, This indicates the minimum transmission hold-up index allowed by the target channel. This represents the maximum allowable additional insertion loss during boundary reconstruction. This represents the maximum number of boundary cells that are allowed to switch states during a single control process. By introducing the above physical constraints, we can avoid the problem of reinforcement learning excessively pursuing crosstalk reduction when searching for boundary encoded states, which could lead to significant degradation of the target channel, excessive additional losses, or frequent switching of boundary cells.

[0023] The described physics-guided deep reinforcement learning network can first generate multi-port RF response data under different operating frequencies, chip combinations, port activation states, and boundary coding conditions based on a full-wave electromagnetic model, an equivalent circuit model, or a combination of both. Offline training can then be completed using this data. Subsequently, a small amount of packaged test data can be used to calibrate the network parameters or the boundary state response model. During online operation, network inference is only required based on the real-time acquired multi-port modal state vectors, and a new boundary coding matrix is ​​output. This does not require re-executing the entire training process. When the average reward change over multiple consecutive training rounds satisfies... Alternatively, when the worst-case port isolation, non-target mode energy, and target channel preservation index reach a preset range, the reinforcement learning control strategy is considered to have reached convergence. and These represent the average reward between two adjacent training phases. This indicates the preset convergence threshold.

[0024] The RF response, modal states, and statistical performance data in this embodiment are obtained based on the constructed packaged electromagnetic response model, theoretical simulation, and finite sample statistical conditions. They are used to illustrate the relationship between multi-port coupling, packaged modal participation, and target channel performance under different boundary control states. Figure 1The overall structure of the microwave RF chip multiport crosstalk self-suppression device and the propagation of RF signals between different chips, interconnect channels, and programmable cavity boundaries within the package are clearly and naturally observed. The upper part shows the multiport microwave excitation, modal state reconstruction, and physically guided deep reinforcement learning control path; the lower part shows the package body composed of RF chips, interposers, RDL, microbumps, a metal cover plate, and a programmable boundary impedance array. From this, it can be clearly concluded that different chip activation combinations and operating frequency bands can excite varying degrees of cavity intrinsic modes, spatial leakage coupling, and multiport crosstalk within the package. Furthermore, by adjusting the equivalent surface impedance of boundary units at different locations, the reflection phase, multipath propagation path, and local field distribution within the package can be actively altered. More importantly, when the dual channels operate in parallel and significant non-target coupling is induced, the non-target coupling of the disturbed port under fixed boundary conditions is about -17.2 dB, while it can be reduced to about -29.6 dB after performing reinforcement learning boundary coding. Meanwhile, the insertion loss of the target RF channel changes only from about 1.83 dB to about 1.96 dB. Therefore, a significant active suppression effect of multi-port crosstalk can be obtained while basically maintaining the normal transmission of the target channel.

[0025] The microwave RF chip packaging device is reasonably mapped to a multi-port modal state space, and then the target transmission response, non-target coupling response, packaging mode participation coefficient, and group delay characteristics under different boundary coding conditions are constructed. (See...) Figure 2 The figure clearly shows significant differences in multiple electromagnetic parameters between different boundary states. For example, near 32.0 GHz, the strongest non-target coupling in the initial fixed boundary state is approximately -18.1 dB, which decreases to approximately -24.7 dB after the first-stage boundary reconstruction, and further decreases to approximately -31.3 dB after physical-guided deep reinforcement learning optimization. Meanwhile, the normalized participation coefficients of the corresponding dominant package cavity modes are approximately 0.78, 0.42, and 0.19, respectively. More importantly, by introducing target channel holding constraints, multi-port coupling constraints, and modal energy constraints, the programmable boundary does not simply absorb all electromagnetic energy, but rather actively cancels out the non-target coupling field near the disturbed port by adjusting the amplitude and phase of different propagation paths. Therefore, its multi-port crosstalk self-suppression characteristic based on the programmable package cavity boundary is clearly demonstrated.

[0026] from Figure 3The distribution of different operational tasks in the multi-port modal state space and the results of physically guided deep reinforcement learning control can be observed very naturally and clearly. The horizontal axis represents the normalized non-target coupling index, and the vertical axis represents the dominant package mode participation coefficient. Simultaneously, a corresponding package operational state vector is formed based on the current operating frequency band, chip activation combination, port operating state, and original boundary coding. This makes it easy to identify that different operational tasks exhibit significantly different modal excitation and crosstalk distributions in their initial states. The single-RF channel operating state mainly clusters in the low-coupling, low-modal participation region, while the multi-chip collaborative operating state is more distributed in the high-coupling, high-modal participation region. In stark contrast, the dual-channel parallel operating state's initial distribution lies between the two. More importantly, by continuously adjusting the coding combination of different boundary impedance units according to the current state through physically guided deep reinforcement learning, different operational tasks can gradually converge to the low non-target coupling and low dominant mode participation region, forming a stable control strategy while satisfying preset constraints on the target channel transmission gain, maximum number of switching units, and additional insertion loss. Therefore, it is natural to construct a self-suppression method for multi-port crosstalk in microwave RF chips based on programmable package cavity boundaries and physical-guided deep reinforcement learning. Thus, the conclusion of this embodiment is very clear: the designed intelligent microwave RF packaging device achieves dynamic reconstruction of the package electromagnetic boundary under different chip combinations and different operating frequency bands, and can effectively reduce multi-port crosstalk while maintaining the performance of the target RF channel through active cancellation of non-target package modes and multipath coupling fields.

[0027] Example 2: In a microwave RF chip multi-port crosstalk self-suppression device and method based on programmable package cavity boundary and physical-guided deep reinforcement learning, two packaging electromagnetic control schemes, namely control device A and control device B, are set up respectively, and their multi-port RF response, package mode suppression capability, and target channel retention capability are compared and analyzed. The device is integrated into a multi-core microwave RF packaging and programmable electromagnetic boundary integrated platform, and the dielectric constant of the packaging interlayer is... The substrate thickness is approximately 0.508 mm, the size of a single RF chip is approximately 5 mm × 5 mm, the chip spacing is approximately 3.0–8.0 mm, the operating frequency band is 20–50 GHz, the preferred analysis frequency band is 28–40 GHz, the programmable package boundary is set with 64 basic impedance units, and the unit state switching error is controlled within ±5%. Figure 4As shown, control device A and control device B, under the same RF chip, package structure, port activation combination, and operating frequency band, can achieve multi-port RF response reconfiguration simply by changing the package electromagnetic boundary control strategy. Control device A uses a fixed metal boundary or a fixed periodic electromagnetic suppression structure and does not change its own parameters according to the current port coupling and package mode state; control device B, on the other hand, uses multi-region programmable package boundary, mode state reconfiguration, and physically guided deep reinforcement learning control. Figure 4 The solid line represents the multi-port RF response before boundary reconstruction, and the dashed line represents the response result after the corresponding electromagnetic control is executed. It can be observed that control device A can only achieve limited reduction of non-target coupling within the preset frequency band, while control device B can regenerate the boundary coding matrix according to the current chip combination and operating frequency band, significantly reducing the peak value of non-target coupling and keeping the target RF channel within the preset transmission performance range, demonstrating the device's dynamic electromagnetic boundary reconstruction capability under different package operation tasks.

[0028] This embodiment rationally maps different packaging control results to a finite-sample multi-port modal state-space model. First, it extracts parameters such as worst-case port isolation, dominant mode participation coefficient, target channel insertion loss, and group delay ripple under different chip combinations, port activation states, and operating frequencies. Then, it uses strategy returns and control convergence indices to naturally and clearly identify the distribution of each state. See details... Figure 5 Specifically, Figure 5 The left side shows the multi-port state distribution under control device A, and the right side shows the results under control device B. The color intensity reflects the overall control benefits under the corresponding package operating conditions. It is clearly visible from the figure that, under fixed boundary conditions, some dual-channel and multi-chip collaborative operating states are still concentrated in regions with higher non-target coupling and higher package mode participation. More importantly, compared to control device A, control device B, which employs programmable cavity boundaries and physically guided deep reinforcement learning, enables most operating states to migrate to regions with lower coupling and lower mode participation. Device A has a worst-case port isolation of approximately 18.6 dB, an average dominant mode suppression rate of approximately 41.0%, and a target channel performance retention rate of approximately 91.0%. In contrast, device B improves the worst-case port isolation to approximately 30.8 dB, the average dominant mode suppression rate to approximately 78.0%, and the target channel performance retention rate to approximately 96.0%. Therefore, adaptive control of multi-port crosstalk under different chip, frequency band, and port task conditions can be achieved through dynamic adjustment of boundary impedance states and reinforcement learning strategies.

[0029] To further quantitatively evaluate the performance of different packaging boundary control schemes, this embodiment defines the worst-case port isolation. Dominant mode suppression rate and target channel performance retention rate Three performance metrics. Worst-case port isolation for all currently active non-target port combinations. Taking the minimum isolation value among all non-target coupling channels, it can be expressed as: in, This represents the set of all non-target port combinations in the current running state. This represents the isolation degree between the i-th port and the j-th port. It is used to characterize the worst crosstalk state in the entire multi-port encapsulation system. The larger the value, the better the isolation performance of the worst non-target coupling channel in the system.

[0030] The dominant mode suppression rate Defined as in, This represents the normalized energy or mode participation index of the dominant non-target encapsulation mode within the target frequency band before boundary reconstruction. This indicates the normalized energy or modal participation index after boundary reconstruction. The higher the value, the more significant the suppression of the dominant non-target encapsulation mode after boundary reconstruction; when... When the boundary is reconstructed, it indicates that the dominant mode has not changed significantly before and after the boundary reconstruction.

[0031] The target channel performance retention rate Defined as Where, represents the total number of frequency sampling points within the target analysis frequency band, Represents the i-th frequency sampling point. This represents the target channel transmission response corresponding to the i-th frequency sampling point after boundary reconstruction. This represents the transmission response of the reference target channel at the same frequency sampling point. Because... Expressed in dB, therefore using After converting it to a linear power ratio, a comprehensive evaluation is performed over the frequency band. When the calculated result is greater than 100%, it can be limited to 100% according to the evaluation requirements. The closer it is to 100%, the smaller the additional impact of boundary reconstruction on normal target radio frequency transmission.

[0032] In this embodiment, the three indicators evaluate different boundary control schemes from three aspects: the most unfavorable port crosstalk state, the degree of suppression of the dominant non-target packaging mode, and the target RF channel retention capability. This avoids the problem of insufficient information dimension caused by using only a single port isolation degree or a single transmission parameter to evaluate the packaging electromagnetic boundary reconstruction effect.

[0033] This embodiment naturally and reasonably introduces random disturbances such as packaging processing dimensional deviations, dielectric parameter fluctuations, boundary device state errors, temperature drift, and measurement noise based on existing foundations. The comprehensive disturbance range is set to ±10% to simulate the manufacturing differences and operational instability factors in actual microwave RF chip packaging. Under the constructed theoretical simulation and finite sample statistical conditions, the basic trend of multi-port crosstalk control performance changing with the comprehensive disturbance ratio is analyzed. (See [link to relevant documentation]). Figure 6 The horizontal axis represents the overall disturbance ratio, and the vertical axis represents the worst-case port isolation, dominant package mode suppression rate, and target channel performance retention rate. Under the constructed theoretical simulation and finite sample statistical conditions, the results show that under a certain range of disturbance conditions, the crosstalk suppression performance of the programmable package boundary is basically stable. When the overall disturbance ratio increases from 0% to 10%, the worst-case port isolation of control device A decreases from 18.6 dB to approximately 15.2 dB, the dominant mode suppression rate decreases from 41.0% to approximately 28.0%, and the target channel performance retention rate decreases from 91.0% to approximately 82.0%. In contrast, the worst-case port isolation of control device B decreases only from 30.8 dB to approximately 27.4 dB, the dominant mode suppression rate decreases from 78.0% to approximately 69.0%, and the target channel performance retention rate remains above approximately 92.0%. This indicates that the programmable package boundary, modal state reconstruction, and physically guided reinforcement learning control structure have stronger tolerance to external disturbances. Therefore, it is quite natural to conclude that the proposed microwave RF chip multi-port crosstalk self-suppression device can reliably maintain the characteristics of package cavity mode suppression and active cancellation of non-target coupling even under the presence of processing errors, temperature changes, device state deviations and measurement noise. In other words, its structure has good robustness in microwave millimeter-wave multi-chip packaging and Chiplet heterogeneous integration applications.

Claims

1. A self-suppression device for multi-port crosstalk of microwave radio frequency chips based on programmable package cavity boundaries and physically guided deep reinforcement learning, characterized in that: The device comprises a microwave RF multi-core package body, a multi-port microwave excitation and signal acquisition module, a programmable package electromagnetic boundary array, a multi-port modal state reconstruction module, a physically guided deep reinforcement learning control module, a boundary coding matrix output module, and a programmable boundary unit driving module. The programmable package electromagnetic boundary array is positioned at at least two different locations on the inner surface of the metal package cover, the package sidewall, the periphery of the interposer layer, the periphery of the redistribution layer, and the non-signal region between the cores, and includes multiple programmable boundary units with at least two equivalent surface impedance states. The multi-port modal state reconstruction module constructs a multi-port modal state vector based on the target channel transmission response, the non-target port coupling response, the package modal participation coefficient, the operating frequency band, the port activation state, and the current boundary coding state. The physically guided deep reinforcement learning control module takes the multi-port modal state vector as input and outputs a new boundary coding matrix, thereby driving the programmable boundary units at different locations to change the local equivalent surface impedance, reflection phase, and multipath propagation conditions, so that the non-target coupling field or the dominant package cavity mode forms a destructive reconstruction at the disturbed port, while simultaneously ensuring that the target RF channel meets preset transmission performance requirements.

2. The microwave RF chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical-guided deep reinforcement learning according to claim 1, characterized in that: Each programmable boundary unit is composed of one or more of the following: switchable inductor, capacitor, PIN diode, variable capacitor, RF switch, RF MEMS, ferroelectric device, phase change device, or other adjustable impedance structure. The equivalent surface impedance, reflection phase, local loss, or operating state of the boundary unit is set according to the target operating frequency band and the electromagnetic field distribution inside the package. Each programmable boundary unit forms a two-dimensional or three-dimensional boundary coding array along the inner surface of the metal cover, the package sidewall, the periphery of the interposer, or the inter-chip region. Different boundary regions are used to adjust the package cavity mode, inter-chip space leakage coupling, multi-port direct crosstalk, and non-target multipath propagation path, respectively.

3. The microwave RF chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical-guided deep reinforcement learning according to claim 1, characterized in that: Multi-parameter feature extraction is performed on the target channel transmission signal, non-target port coupling signal, and packaged modal response obtained during multi-port microwave excitation and signal acquisition. This forms a multi-port modal state vector containing at least two or more types of parameters, including operating frequency, port activation matrix, non-target coupling matrix, dominant packaged modal participation coefficient, target channel transmission performance, and current boundary coding state. By constructing a multi-port modal state space, the mapping relationship between packaged operation tasks, electromagnetic propagation state inside the package, and boundary coding matrix is ​​realized, thereby forming a reconfigurable electromagnetic boundary control structure for different frequency bands, different core combinations, and different port tasks.

4. The microwave RF chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical-guided deep reinforcement learning according to claim 1, characterized in that: The physical-guided deep reinforcement learning control module forms different electromagnetic boundary reconstruction states in the packaging device by changing the weight ratios between different modal state parameters, boundary coding strategies, reward function parameters, and physical constraint parameters. When non-target port coupling features dominate, the first type of boundary coding matrix is ​​preferentially output to improve the worst-case port isolation. When the dominant packaging cavity modal participation coefficient dominates, the second type of boundary coding matrix is ​​preferentially output to reduce the energy of non-target modes. When multi-port coupling, modal participation, target channel preservation, and group delay features work together, the device presents a multi-parameter joint electromagnetic boundary reconstruction state. The physical-guided deep reinforcement learning control module further introduces one or more of the following constraints: minimum transmission constraint of target channel, worst-case port isolation constraint, non-target mode energy constraint, additional insertion loss constraint, boundary unit switching quantity constraint, device bias range constraint, and packaging temperature rise constraint, so that the boundary reconstruction results are robust under conditions of temperature drift, manufacturing error, dielectric parameter fluctuation, boundary device state deviation, and measurement noise.

5. The microwave RF chip multi-port crosstalk self-suppression device based on programmable package cavity boundary and physical-guided deep reinforcement learning according to claim 4, characterized in that: The multi-port microwave excitation and signal acquisition module and the physically guided deep reinforcement learning control module are used to change the operating frequency band, chip activation combination, port operating state, modal state extraction method, boundary impedance state, and boundary coding matrix, thereby controlling the opening and closing of different package electromagnetic boundary regions and crosstalk suppression channels. When the multi-port modal state indicates that non-target coupling or the dominant package cavity mode exceeds the preset conditions, the programmable boundary unit at the corresponding position is driven to perform impedance state switching and form a destructive reconstruction path. When the target channel transmission performance is lower than the preset threshold or the boundary control does not meet the physical constraints, the corresponding boundary coding action is stopped and the system is restored to the previous effective state or the preset safe state. When the multi-port coupling, modal participation, and target channel holding conditions meet the set requirements, a package control result with stable non-target coupling suppression, dominant mode reduction, target channel holding, and dynamic electromagnetic boundary reconstruction capabilities is formed.

6. A method for self-suppression of multi-port crosstalk in a microwave radio frequency chip using the apparatus described in any one of claims 1 to 5, characterized in that, Includes the following steps: S1, the target channel transmission response, non-target port coupling response and encapsulation mode response are obtained through the multi-port microwave excitation and signal acquisition module; S2, construct a multi-port mode state vector based on the operating frequency, port activation state, non-target coupling matrix, dominant encapsulation mode participation coefficient, target channel transmission performance and current boundary coding state; S3, input the multi-port modal state vector into the physical-guided deep reinforcement learning control module to obtain a new boundary encoding matrix; S4, drive the programmable boundary units at different positions to change the equivalent surface impedance, reflection phase or on / off state according to the boundary coding matrix; S5 enables the non-target coupling field or the dominant package cavity mode to form a destructive reconstruction at the disturbed port, while determining whether the target RF channel meets the preset transmission performance requirements; when the preset physical constraints are not met, it restores to the previous effective boundary state or the preset safe state.

Images